Save 37% off PRO during our Black Friday Sale! »

LPixelLT20191121.pdf

7eafc7c18c2652c86e0b5f242f2d5e58?s=47 Tsuyama
November 21, 2019
230

 LPixelLT20191121.pdf

7eafc7c18c2652c86e0b5f242f2d5e58?s=128

Tsuyama

November 21, 2019
Tweet

Transcript

  1. Context-Aware Convolutional Neural Network for Grading of Colorectal Cancer Histology

    Images h"ps://arxiv.org/abs/1907.09478 (Submi'ed on 22 Jul 2019) Muhammad Shaban, Ruqayya Awan, Muhammad Moazam Fraz, Senior Member, IEEE, Ayesha Azam, David Snead *Nasir M. Rajpoot, Senior Member, IEEE Department of Computer Science, University of Warwick LPixel Inc. Presents Image Analysis x Machine Learning #8 21th Nov. 2019 @LPixel Inc. 病理医 Tu-chan
  2. § デジタル化されたスライドガラス(whole-slide image: WSI)はファイルサイズが巨⼤なので計算コスト の制約から⼩さなパッチ(タイル)に分割し,さらにダウンサンプルして学習に⽤いる⼿法が⼀般 的だが,オリジナル画像が持つ”context”が失われてしまう 背景 2 Sirinukunwattana et

    al. MICCAI’2018 A cat A cat is jumping A cat is jumping catching a butterfly Benign Tumour 20x 5x 10x 2.5x Fig. 1. Visual context. The different images of the scene containing a jumping cat effectively highlight that the correct interpretation of a scene depends on visual context. We content the accuracy of dense segmentation of histology images into different tissue types depends on our ability to make effective use of multiple scales. context directly from training data. With this paper we provide a more system- atic comparison of these approaches and study how these effect the ability of differentiating between different tissue components. In addition, we introduce a computational model that utilises feature sharing across scales and learns de- pendencies between scales using long-short term memory (LSTM) unit [5]. An openly available collection of breast cancer samples [6] and a local col- lection of prostate cancer histology provide the necessary disease context. An overview of the relevant deep learning approaches is provided in Section 2. The 切り出したりダウンサンプルすると診断に必要な全体構造や細胞レベルの情報が失われる RWALLA, SHABAN, RAJPOOT: REPRESENTATION-AGGREGATION NETWORKS 25 m 100 m 400 m 800 m 10 mm ure 1: A whole slide image and multi-scale visualization of a sub region. https://arxiv.org/abs/1707.08814 WSI @40x 200K x 100K pixel 0.25 micron/pixel https://arxiv.org/abs/1806.04259 § 病理診断プロセスは⾼倍率から低倍率にむかって, 組織全体の構築→細胞レベルの形態学の観察が必要 § 局所の特徴と全体構築を⼀緒に学習するにはどうしたら よいか?
  3. h"ps://arxiv.org/abs/1806.04259 Context-aware Learning: 過去の研究 Improving Whole Slide Segmentation Through Visual

    Context 3 64 64 px <latexit sha1_base64="wPNkypMEmF0JnyuEj3hqRDozLUA=">AAAB/XicbVDLSsNAFJ34rPUVFVduBovgQkoixceu4MZlBWMLTSiT6aQdOpmEmRtpCQV/xY0LFbf+hzv/xmmbhbYeGOZwzr3ce0+YCq7Bcb6tpeWV1bX10kZ5c2t7Z9fe23/QSaYo82giEtUKiWaCS+YBB8FaqWIkDgVrhoObid98ZErzRN7DKGVBTHqSR5wSMFLHPryo+cBjps1/5gMbQp4Oxx274lSdKfAicQtSQQUaHfvL7yY0i5kEKojWbddJIciJAk4FG5f9TLOU0AHpsbahkpiJQT5df4xPjNLFUaLMk4Cn6u+OnMRaj+LQVMYE+nrem4j/ee0Moqsg5zLNgEk6GxRlAkOCJ1ngLleMghgZQqjiZldM+0QRCiaxsgnBnT95kXjn1euqc1er1GtFGiV0hI7RKXLRJaqjW9RAHqIoR8/oFb1ZT9aL9W59zEqXrKLnAP2B9fkDHvyVHQ==</latexit> <latexit sha1_base64="wPNkypMEmF0JnyuEj3hqRDozLUA=">AAAB/XicbVDLSsNAFJ34rPUVFVduBovgQkoixceu4MZlBWMLTSiT6aQdOpmEmRtpCQV/xY0LFbf+hzv/xmmbhbYeGOZwzr3ce0+YCq7Bcb6tpeWV1bX10kZ5c2t7Z9fe23/QSaYo82giEtUKiWaCS+YBB8FaqWIkDgVrhoObid98ZErzRN7DKGVBTHqSR5wSMFLHPryo+cBjps1/5gMbQp4Oxx274lSdKfAicQtSQQUaHfvL7yY0i5kEKojWbddJIciJAk4FG5f9TLOU0AHpsbahkpiJQT5df4xPjNLFUaLMk4Cn6u+OnMRaj+LQVMYE+nrem4j/ee0Moqsg5zLNgEk6GxRlAkOCJ1ngLleMghgZQqjiZldM+0QRCiaxsgnBnT95kXjn1euqc1er1GtFGiV0hI7RKXLRJaqjW9RAHqIoR8/oFb1ZT9aL9W59zEqXrKLnAP2B9fkDHvyVHQ==</latexit> <latexit sha1_base64="wPNkypMEmF0JnyuEj3hqRDozLUA=">AAAB/XicbVDLSsNAFJ34rPUVFVduBovgQkoixceu4MZlBWMLTSiT6aQdOpmEmRtpCQV/xY0LFbf+hzv/xmmbhbYeGOZwzr3ce0+YCq7Bcb6tpeWV1bX10kZ5c2t7Z9fe23/QSaYo82giEtUKiWaCS+YBB8FaqWIkDgVrhoObid98ZErzRN7DKGVBTHqSR5wSMFLHPryo+cBjps1/5gMbQp4Oxx274lSdKfAicQtSQQUaHfvL7yY0i5kEKojWbddJIciJAk4FG5f9TLOU0AHpsbahkpiJQT5df4xPjNLFUaLMk4Cn6u+OnMRaj+LQVMYE+nrem4j/ee0Moqsg5zLNgEk6GxRlAkOCJ1ngLleMghgZQqjiZldM+0QRCiaxsgnBnT95kXjn1euqc1er1GtFGiV0hI7RKXLRJaqjW9RAHqIoR8/oFb1ZT9aL9W59zEqXrKLnAP2B9fkDHvyVHQ==</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> 20 <latexit sha1_base64="u5+bhCm50+8j9uA/ZM6vtP5DiUo=">AAAB7nicbVBNS8NAEJ34WetX1aOXxSJ4KkkpqLeCF48VjC20oWy2m3bpZhN3J0IJ/RNePKh49fd489+4bXPQ1gcDj/dmmJkXplIYdN1vZ219Y3Nru7RT3t3bPzisHB0/mCTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8PxzcxvP3FtRKLucZLyIKZDJSLBKFqpU3d7KGJu+pWqW3PnIKvEK0gVCrT6la/eIGFZzBUySY3pem6KQU41Cib5tNzLDE8pG9Mh71qqqF0S5PN7p+TcKgMSJdqWQjJXf0/kNDZmEoe2M6Y4MsveTPzP62YYXQW5UGmGXLHFoiiTBBMye54MhOYM5cQSyrSwtxI2opoytBGVbQje8surxK/XrmvuXaPabBRplOAUzuACPLiEJtxCC3xgIOEZXuHNeXRenHfnY9G65hQzJ/AHzucP/n6Pbw==</latexit> <latexit sha1_base64="u5+bhCm50+8j9uA/ZM6vtP5DiUo=">AAAB7nicbVBNS8NAEJ34WetX1aOXxSJ4KkkpqLeCF48VjC20oWy2m3bpZhN3J0IJ/RNePKh49fd489+4bXPQ1gcDj/dmmJkXplIYdN1vZ219Y3Nru7RT3t3bPzisHB0/mCTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8PxzcxvP3FtRKLucZLyIKZDJSLBKFqpU3d7KGJu+pWqW3PnIKvEK0gVCrT6la/eIGFZzBUySY3pem6KQU41Cib5tNzLDE8pG9Mh71qqqF0S5PN7p+TcKgMSJdqWQjJXf0/kNDZmEoe2M6Y4MsveTPzP62YYXQW5UGmGXLHFoiiTBBMye54MhOYM5cQSyrSwtxI2opoytBGVbQje8surxK/XrmvuXaPabBRplOAUzuACPLiEJtxCC3xgIOEZXuHNeXRenHfnY9G65hQzJ/AHzucP/n6Pbw==</latexit> <latexit sha1_base64="u5+bhCm50+8j9uA/ZM6vtP5DiUo=">AAAB7nicbVBNS8NAEJ34WetX1aOXxSJ4KkkpqLeCF48VjC20oWy2m3bpZhN3J0IJ/RNePKh49fd489+4bXPQ1gcDj/dmmJkXplIYdN1vZ219Y3Nru7RT3t3bPzisHB0/mCTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8PxzcxvP3FtRKLucZLyIKZDJSLBKFqpU3d7KGJu+pWqW3PnIKvEK0gVCrT6la/eIGFZzBUySY3pem6KQU41Cib5tNzLDE8pG9Mh71qqqF0S5PN7p+TcKgMSJdqWQjJXf0/kNDZmEoe2M6Y4MsveTPzP62YYXQW5UGmGXLHFoiiTBBMye54MhOYM5cQSyrSwtxI2opoytBGVbQje8surxK/XrmvuXaPabBRplOAUzuACPLiEJtxCC3xgIOEZXuHNeXRenHfnY9G65hQzJ/AHzucP/n6Pbw==</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> 10 <latexit sha1_base64="4gnuXD/DK6fCiDxEIDlhwz41Zps=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoN4KXjxWMLbQhrLZbtqlm03cnQgl9E948aDi1d/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTx6MEmmGfdZIhPdCanhUijuo0DJO6nmNA4lb4fjm5nffuLaiETd4yTlQUyHSkSCUbRSx3N7KGJu+tWaW3fnIKvEK0gNCrT61a/eIGFZzBUySY3pem6KQU41Cib5tNLLDE8pG9Mh71qqqF0S5PN7p+TMKgMSJdqWQjJXf0/kNDZmEoe2M6Y4MsveTPzP62YYXQW5UGmGXLHFoiiTBBMye54MhOYM5cQSyrSwtxI2opoytBFVbAje8surxL+oX9fdu0at2SjSKMMJnMI5eHAJTbiFFvjAQMIzvMKb8+i8OO/Ox6K15BQzx/AHzucP/PSPbg==</latexit> <latexit sha1_base64="4gnuXD/DK6fCiDxEIDlhwz41Zps=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoN4KXjxWMLbQhrLZbtqlm03cnQgl9E948aDi1d/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTx6MEmmGfdZIhPdCanhUijuo0DJO6nmNA4lb4fjm5nffuLaiETd4yTlQUyHSkSCUbRSx3N7KGJu+tWaW3fnIKvEK0gNCrT61a/eIGFZzBUySY3pem6KQU41Cib5tNLLDE8pG9Mh71qqqF0S5PN7p+TMKgMSJdqWQjJXf0/kNDZmEoe2M6Y4MsveTPzP62YYXQW5UGmGXLHFoiiTBBMye54MhOYM5cQSyrSwtxI2opoytBFVbAje8surxL+oX9fdu0at2SjSKMMJnMI5eHAJTbiFFvjAQMIzvMKb8+i8OO/Ox6K15BQzx/AHzucP/PSPbg==</latexit> <latexit sha1_base64="4gnuXD/DK6fCiDxEIDlhwz41Zps=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoN4KXjxWMLbQhrLZbtqlm03cnQgl9E948aDi1d/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTx6MEmmGfdZIhPdCanhUijuo0DJO6nmNA4lb4fjm5nffuLaiETd4yTlQUyHSkSCUbRSx3N7KGJu+tWaW3fnIKvEK0gNCrT61a/eIGFZzBUySY3pem6KQU41Cib5tNLLDE8pG9Mh71qqqF0S5PN7p+TMKgMSJdqWQjJXf0/kNDZmEoe2M6Y4MsveTPzP62YYXQW5UGmGXLHFoiiTBBMye54MhOYM5cQSyrSwtxI2opoytBFVbAje8surxL+oX9fdu0at2SjSKMMJnMI5eHAJTbiFFvjAQMIzvMKb8+i8OO/Ox6K15BQzx/AHzucP/PSPbg==</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> 5 <latexit sha1_base64="iliZc0WaPwjRXhG52I3HRPU1+fk=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkot4KXjxWMLbQhrLZbtqlm03YnQgl9Ed48aDi1f/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTx6NEmmGfdZIhPdCanhUijuo0DJO6nmNA4lb4fj25nffuLaiEQ94CTlQUyHSkSCUbRS+7KHIuamX625dXcOskq8gtSgQKtf/eoNEpbFXCGT1Jiu56YY5FSjYJJPK73M8JSyMR3yrqWK2iVBPj93Ss6sMiBRom0pJHP190ROY2MmcWg7Y4ojs+zNxP+8bobRdZALlWbIFVssijJJMCGz38lAaM5QTiyhTAt7K2EjqilDm1DFhuAtv7xK/Iv6Td29b9SajSKNMpzAKZyDB1fQhDtogQ8MxvAMr/DmpM6L8+58LFpLTjFzDH/gfP4AlEiPOA==</latexit> <latexit sha1_base64="iliZc0WaPwjRXhG52I3HRPU1+fk=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkot4KXjxWMLbQhrLZbtqlm03YnQgl9Ed48aDi1f/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTx6NEmmGfdZIhPdCanhUijuo0DJO6nmNA4lb4fj25nffuLaiEQ94CTlQUyHSkSCUbRS+7KHIuamX625dXcOskq8gtSgQKtf/eoNEpbFXCGT1Jiu56YY5FSjYJJPK73M8JSyMR3yrqWK2iVBPj93Ss6sMiBRom0pJHP190ROY2MmcWg7Y4ojs+zNxP+8bobRdZALlWbIFVssijJJMCGz38lAaM5QTiyhTAt7K2EjqilDm1DFhuAtv7xK/Iv6Td29b9SajSKNMpzAKZyDB1fQhDtogQ8MxvAMr/DmpM6L8+58LFpLTjFzDH/gfP4AlEiPOA==</latexit> <latexit sha1_base64="iliZc0WaPwjRXhG52I3HRPU1+fk=">AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkot4KXjxWMLbQhrLZbtqlm03YnQgl9Ed48aDi1f/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTx6NEmmGfdZIhPdCanhUijuo0DJO6nmNA4lb4fj25nffuLaiEQ94CTlQUyHSkSCUbRS+7KHIuamX625dXcOskq8gtSgQKtf/eoNEpbFXCGT1Jiu56YY5FSjYJJPK73M8JSyMR3yrqWK2iVBPj93Ss6sMiBRom0pJHP190ROY2MmcWg7Y4ojs+zNxP+8bobRdZALlWbIFVssijJJMCGz38lAaM5QTiyhTAt7K2EjqilDm1DFhuAtv7xK/Iv6Td29b9SajSKNMpzAKZyDB1fQhDtogQ8MxvAMr/DmpM6L8+58LFpLTjFzDH/gfP4AlEiPOA==</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> 2.5 <latexit sha1_base64="5mf2lDhnG2o3KP9pVROCPpFPCuw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0hKRb0VvHisYGylDWWz3bRLdzdhdyOU0F/hxYOKV/+ON/+N2zYHbX0w8Hhvhpl5UcqZNp737ZTW1jc2t8rblZ3dvf2D6uHRg04yRWhAEp6oToQ15UzSwDDDaSdVFIuI03Y0vpn57SeqNEvkvZmkNBR4KFnMCDZWeqy7Fz3DBNX9as1zvTnQKvELUoMCrX71qzdISCaoNIRjrbu+l5owx8owwum00ss0TTEZ4yHtWiqxXRLm84On6MwqAxQnypY0aK7+nsix0HoiItspsBnpZW8m/ud1MxNfhTmTaWaoJItFccaRSdDsezRgihLDJ5Zgopi9FZERVpgYm1HFhuAvv7xKgrp77Xp3jVqzUaRRhhM4hXPw4RKacAstCICAgGd4hTdHOS/Ou/OxaC05xcwx/IHz+QNyKY+s</latexit> <latexit sha1_base64="5mf2lDhnG2o3KP9pVROCPpFPCuw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0hKRb0VvHisYGylDWWz3bRLdzdhdyOU0F/hxYOKV/+ON/+N2zYHbX0w8Hhvhpl5UcqZNp737ZTW1jc2t8rblZ3dvf2D6uHRg04yRWhAEp6oToQ15UzSwDDDaSdVFIuI03Y0vpn57SeqNEvkvZmkNBR4KFnMCDZWeqy7Fz3DBNX9as1zvTnQKvELUoMCrX71qzdISCaoNIRjrbu+l5owx8owwum00ss0TTEZ4yHtWiqxXRLm84On6MwqAxQnypY0aK7+nsix0HoiItspsBnpZW8m/ud1MxNfhTmTaWaoJItFccaRSdDsezRgihLDJ5Zgopi9FZERVpgYm1HFhuAvv7xKgrp77Xp3jVqzUaRRhhM4hXPw4RKacAstCICAgGd4hTdHOS/Ou/OxaC05xcwx/IHz+QNyKY+s</latexit> <latexit sha1_base64="5mf2lDhnG2o3KP9pVROCPpFPCuw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0hKRb0VvHisYGylDWWz3bRLdzdhdyOU0F/hxYOKV/+ON/+N2zYHbX0w8Hhvhpl5UcqZNp737ZTW1jc2t8rblZ3dvf2D6uHRg04yRWhAEp6oToQ15UzSwDDDaSdVFIuI03Y0vpn57SeqNEvkvZmkNBR4KFnMCDZWeqy7Fz3DBNX9as1zvTnQKvELUoMCrX71qzdISCaoNIRjrbu+l5owx8owwum00ss0TTEZ4yHtWiqxXRLm84On6MwqAxQnypY0aK7+nsix0HoiItspsBnpZW8m/ud1MxNfhTmTaWaoJItFccaRSdDsezRgihLDJ5Zgopi9FZERVpgYm1HFhuAvv7xKgrp77Xp3jVqzUaRRhhM4hXPw4RKacAstCICAgGd4hTdHOS/Ou/OxaC05xcwx/IHz+QNyKY+s</latexit> A B C D Single scale Cat <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> E Early fusion Cat <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN module <latexit sha1_base64="7V2w3jgS+nWM1jclmruHlM1BCYA=">AAAB+XicbVBNT8JAEN3iF+JX0aOXjcTEEymGRL2RcPFEMLFCAg3ZbrewYbdtdqcqqfwULx7UePWfePPfuEAPCr5kkpf3ZnZnnp8IrsFxvq3C2vrG5lZxu7Szu7d/YJcP73ScKspcGotYdX2imeARc4GDYN1EMSJ9wTr+uDnzO/dMaR5HtzBJmCfJMOIhpwSMNLDLfWCPkDVbLSzjIBVsOrArTtWZA6+SWk4qKEd7YH/1g5imkkVABdG6V3MS8DKigFPzXqmfapYQOiZD1jM0IpJpL5uvPsWnRglwGCtTEeC5+nsiI1LrifRNpyQw0sveTPzP66UQXnoZj5IUWEQXH4WpwBDjWQ444IpREBNDCFXc7IrpiChCwaRVMiHUlk9eJe559arq3NQrjXqeRhEdoxN0hmroAjXQNWojF1H0gJ7RK3qznqwX6936WLQWrHzmCP2B9fkDl6WTuA==</latexit> <latexit sha1_base64="7V2w3jgS+nWM1jclmruHlM1BCYA=">AAAB+XicbVBNT8JAEN3iF+JX0aOXjcTEEymGRL2RcPFEMLFCAg3ZbrewYbdtdqcqqfwULx7UePWfePPfuEAPCr5kkpf3ZnZnnp8IrsFxvq3C2vrG5lZxu7Szu7d/YJcP73ScKspcGotYdX2imeARc4GDYN1EMSJ9wTr+uDnzO/dMaR5HtzBJmCfJMOIhpwSMNLDLfWCPkDVbLSzjIBVsOrArTtWZA6+SWk4qKEd7YH/1g5imkkVABdG6V3MS8DKigFPzXqmfapYQOiZD1jM0IpJpL5uvPsWnRglwGCtTEeC5+nsiI1LrifRNpyQw0sveTPzP66UQXnoZj5IUWEQXH4WpwBDjWQ444IpREBNDCFXc7IrpiChCwaRVMiHUlk9eJe559arq3NQrjXqeRhEdoxN0hmroAjXQNWojF1H0gJ7RK3qznqwX6936WLQWrHzmCP2B9fkDl6WTuA==</latexit> <latexit sha1_base64="7V2w3jgS+nWM1jclmruHlM1BCYA=">AAAB+XicbVBNT8JAEN3iF+JX0aOXjcTEEymGRL2RcPFEMLFCAg3ZbrewYbdtdqcqqfwULx7UePWfePPfuEAPCr5kkpf3ZnZnnp8IrsFxvq3C2vrG5lZxu7Szu7d/YJcP73ScKspcGotYdX2imeARc4GDYN1EMSJ9wTr+uDnzO/dMaR5HtzBJmCfJMOIhpwSMNLDLfWCPkDVbLSzjIBVsOrArTtWZA6+SWk4qKEd7YH/1g5imkkVABdG6V3MS8DKigFPzXqmfapYQOiZD1jM0IpJpL5uvPsWnRglwGCtTEeC5+nsiI1LrifRNpyQw0sveTPzP66UQXnoZj5IUWEQXH4WpwBDjWQ444IpREBNDCFXc7IrpiChCwaRVMiHUlk9eJe559arq3NQrjXqeRhEdoxN0hmroAjXQNWojF1H0gJ7RK3qznqwX6936WLQWrHzmCP2B9fkDl6WTuA==</latexit> Concatenation module <latexit sha1_base64="r0BhPvbbivFfpiGlBNBDQ6Gk5HY=">AAACA3icbVBLSwMxGMzWV62vVY+9BIvgqexKQb0VevFYwbWFtpRsNtuG5rEkWbEsPXjxr3jxoOLVP+HNf2O63YO2DgSGme9LMhMmjGrjed9OaW19Y3OrvF3Z2d3bP3APj+60TBUmAZZMqm6INGFUkMBQw0g3UQTxkJFOOGnN/c49UZpKcWumCRlwNBI0phgZKw3dat+QB5O1pLAKEbkKuYxSRmZDt+bVvRxwlfgFqYEC7aH71Y8kTjkRBjOkdc/3EjPIkDIU2/sq/VSTBOEJGpGepQJxogdZHmIGT60SwVgqe4SBufp7I0Nc6ykP7SRHZqyXvbn4n9dLTXw5yKhIUhsQLx6KUwaNhPNGYEQVwYZNLUFYUftXiMdIIWxsbxVbgr8ceZUE5/WrunfTqDUbRRtlUAUn4Az44AI0wTVogwBg8AiewSt4c56cF+fd+ViMlpxi5xj8gfP5Aw6UmIM=</latexit> <latexit sha1_base64="r0BhPvbbivFfpiGlBNBDQ6Gk5HY=">AAACA3icbVBLSwMxGMzWV62vVY+9BIvgqexKQb0VevFYwbWFtpRsNtuG5rEkWbEsPXjxr3jxoOLVP+HNf2O63YO2DgSGme9LMhMmjGrjed9OaW19Y3OrvF3Z2d3bP3APj+60TBUmAZZMqm6INGFUkMBQw0g3UQTxkJFOOGnN/c49UZpKcWumCRlwNBI0phgZKw3dat+QB5O1pLAKEbkKuYxSRmZDt+bVvRxwlfgFqYEC7aH71Y8kTjkRBjOkdc/3EjPIkDIU2/sq/VSTBOEJGpGepQJxogdZHmIGT60SwVgqe4SBufp7I0Nc6ykP7SRHZqyXvbn4n9dLTXw5yKhIUhsQLx6KUwaNhPNGYEQVwYZNLUFYUftXiMdIIWxsbxVbgr8ceZUE5/WrunfTqDUbRRtlUAUn4Az44AI0wTVogwBg8AiewSt4c56cF+fd+ViMlpxi5xj8gfP5Aw6UmIM=</latexit> <latexit sha1_base64="r0BhPvbbivFfpiGlBNBDQ6Gk5HY=">AAACA3icbVBLSwMxGMzWV62vVY+9BIvgqexKQb0VevFYwbWFtpRsNtuG5rEkWbEsPXjxr3jxoOLVP+HNf2O63YO2DgSGme9LMhMmjGrjed9OaW19Y3OrvF3Z2d3bP3APj+60TBUmAZZMqm6INGFUkMBQw0g3UQTxkJFOOGnN/c49UZpKcWumCRlwNBI0phgZKw3dat+QB5O1pLAKEbkKuYxSRmZDt+bVvRxwlfgFqYEC7aH71Y8kTjkRBjOkdc/3EjPIkDIU2/sq/VSTBOEJGpGepQJxogdZHmIGT60SwVgqe4SBufp7I0Nc6ykP7SRHZqyXvbn4n9dLTXw5yKhIUhsQLx6KUwaNhPNGYEQVwYZNLUFYUftXiMdIIWxsbxVbgr8ceZUE5/WrunfTqDUbRRtlUAUn4Az44AI0wTVogwBg8AiewSt4c56cF+fd+ViMlpxi5xj8gfP5Aw6UmIM=</latexit> LSTM module <latexit sha1_base64="EMicCZKTcrtAvR+n0zk9QGu/0Mw=">AAAB+nicbVBNS8NAEN3Ur1q/Yj16WSyCp5JKQb0VvHhQqNjYQhvKZrNpl242YXciLaF/xYsHFa/+Em/+G7dtDtr6YODx3szuzPMTwTU4zrdVWFvf2Nwqbpd2dvf2D+zD8qOOU0WZS2MRq45PNBNcMhc4CNZJFCORL1jbH13P/PYTU5rHsgWThHkRGUgeckrASH273AM2huz2oXWHozhIBZv27YpTdebAq6SWkwrK0ezbX70gpmnEJFBBtO7WnAS8jCjg1LxX6qWaJYSOyIB1DZUkYtrL5rtP8alRAhzGypQEPFd/T2Qk0noS+aYzIjDUy95M/M/rphBeehmXSQpM0sVHYSowxHgWBA64YhTExBBCFTe7YjokilAwcZVMCLXlk1eJe169qjr39UqjnqdRRMfoBJ2hGrpADXSDmshFFI3RM3pFb9bUerHerY9Fa8HKZ47QH1ifP1YnlCM=</latexit> <latexit sha1_base64="EMicCZKTcrtAvR+n0zk9QGu/0Mw=">AAAB+nicbVBNS8NAEN3Ur1q/Yj16WSyCp5JKQb0VvHhQqNjYQhvKZrNpl242YXciLaF/xYsHFa/+Em/+G7dtDtr6YODx3szuzPMTwTU4zrdVWFvf2Nwqbpd2dvf2D+zD8qOOU0WZS2MRq45PNBNcMhc4CNZJFCORL1jbH13P/PYTU5rHsgWThHkRGUgeckrASH273AM2huz2oXWHozhIBZv27YpTdebAq6SWkwrK0ezbX70gpmnEJFBBtO7WnAS8jCjg1LxX6qWaJYSOyIB1DZUkYtrL5rtP8alRAhzGypQEPFd/T2Qk0noS+aYzIjDUy95M/M/rphBeehmXSQpM0sVHYSowxHgWBA64YhTExBBCFTe7YjokilAwcZVMCLXlk1eJe169qjr39UqjnqdRRMfoBJ2hGrpADXSDmshFFI3RM3pFb9bUerHerY9Fa8HKZ47QH1ifP1YnlCM=</latexit> <latexit sha1_base64="EMicCZKTcrtAvR+n0zk9QGu/0Mw=">AAAB+nicbVBNS8NAEN3Ur1q/Yj16WSyCp5JKQb0VvHhQqNjYQhvKZrNpl242YXciLaF/xYsHFa/+Em/+G7dtDtr6YODx3szuzPMTwTU4zrdVWFvf2Nwqbpd2dvf2D+zD8qOOU0WZS2MRq45PNBNcMhc4CNZJFCORL1jbH13P/PYTU5rHsgWThHkRGUgeckrASH273AM2huz2oXWHozhIBZv27YpTdebAq6SWkwrK0ezbX70gpmnEJFBBtO7WnAS8jCjg1LxX6qWaJYSOyIB1DZUkYtrL5rtP8alRAhzGypQEPFd/T2Qk0noS+aYzIjDUy95M/M/rphBeehmXSQpM0sVHYSowxHgWBA64YhTExBBCFTe7YjokilAwcZVMCLXlk1eJe169qjr39UqjnqdRRMfoBJ2hGrpADXSDmshFFI3RM3pFb9bUerHerY9Fa8HKZ47QH1ifP1YnlCM=</latexit> Prediction module <latexit sha1_base64="TONIb0j9q52JrJP8M/YXNM00QjA=">AAACAHicbVA9SwNBEN2LXzF+RW0Em8UgWIU7CahdwMYygmeEJIS9vblkyd7esTsnhiM2/hUbCxVbf4ad/8bNR6GJDwYe783szrwglcKg6347haXlldW14nppY3Nre6e8u3drkkxz8HkiE30XMANSKPBRoIS7VAOLAwnNYHA59pv3oI1I1A0OU+jErKdEJDhDK3XLB22EB8wbGkLBxxKNkzCTMOqWK27VnYAuEm9GKmSGRrf81Q4TnsWgkEtmTMtzU+zkTKPg9r1SOzOQMj5gPWhZqlgMppNPLhjRY6uENEq0LYV0ov6eyFlszDAObGfMsG/mvbH4n9fKMDrv5EKlGYLi04+iTFJM6DgOGgoNHOXQEsa1sLtS3meacbShlWwI3vzJi8Q/rV5U3etapV6bpVEkh+SInBCPnJE6uSIN4hNOHskzeSVvzpPz4rw7H9PWgjOb2Sd/4Hz+AK1elzA=</latexit> <latexit sha1_base64="TONIb0j9q52JrJP8M/YXNM00QjA=">AAACAHicbVA9SwNBEN2LXzF+RW0Em8UgWIU7CahdwMYygmeEJIS9vblkyd7esTsnhiM2/hUbCxVbf4ad/8bNR6GJDwYe783szrwglcKg6347haXlldW14nppY3Nre6e8u3drkkxz8HkiE30XMANSKPBRoIS7VAOLAwnNYHA59pv3oI1I1A0OU+jErKdEJDhDK3XLB22EB8wbGkLBxxKNkzCTMOqWK27VnYAuEm9GKmSGRrf81Q4TnsWgkEtmTMtzU+zkTKPg9r1SOzOQMj5gPWhZqlgMppNPLhjRY6uENEq0LYV0ov6eyFlszDAObGfMsG/mvbH4n9fKMDrv5EKlGYLi04+iTFJM6DgOGgoNHOXQEsa1sLtS3meacbShlWwI3vzJi8Q/rV5U3etapV6bpVEkh+SInBCPnJE6uSIN4hNOHskzeSVvzpPz4rw7H9PWgjOb2Sd/4Hz+AK1elzA=</latexit> <latexit sha1_base64="TONIb0j9q52JrJP8M/YXNM00QjA=">AAACAHicbVA9SwNBEN2LXzF+RW0Em8UgWIU7CahdwMYygmeEJIS9vblkyd7esTsnhiM2/hUbCxVbf4ad/8bNR6GJDwYe783szrwglcKg6347haXlldW14nppY3Nre6e8u3drkkxz8HkiE30XMANSKPBRoIS7VAOLAwnNYHA59pv3oI1I1A0OU+jErKdEJDhDK3XLB22EB8wbGkLBxxKNkzCTMOqWK27VnYAuEm9GKmSGRrf81Q4TnsWgkEtmTMtzU+zkTKPg9r1SOzOQMj5gPWhZqlgMppNPLhjRY6uENEq0LYV0ov6eyFlszDAObGfMsG/mvbH4n9fKMDrv5EKlGYLi04+iTFJM6DgOGgoNHOXQEsa1sLtS3meacbShlWwI3vzJi8Q/rV5U3etapV6bpVEkh+SInBCPnJE6uSIN4hNOHskzeSVvzpPz4rw7H9PWgjOb2Sd/4Hz+AK1elzA=</latexit> H P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> Single scale original resolution 512 512px <latexit sha1_base64="XVa39VNO1z67Cq/lUyFwYpXmj2k=">AAAB/3icbZDNSsNAFIUn/tb6F3Xhws1gEVyVpFTUXcGNywrGFppQJtNJO3QyCTM30hK68VXcuFBx62u4822ctllo64GBj3Pv5c49YSq4Bsf5tlZW19Y3Nktb5e2d3b19++DwQSeZosyjiUhUOySaCS6ZBxwEa6eKkTgUrBUOb6b11iNTmifyHsYpC2LSlzzilICxuvbxhVvzgcdMY0PYBzaCPB1NunbFqToz4WVwC6igQs2u/eX3EprFTAIVROuO66QQ5EQBp4JNyn6mWUrokPRZx6AkZmWQzw6Y4DPj9HCUKPMk4Jn7eyInsdbjODSdMYGBXqxNzf9qnQyiqyDnMs2ASTpfFGUCQ4KnaeAeV4yCGBsgVHHzV0wHRBEKJrOyCcFdPHkZvFr1uurc1SuNepFGCZ2gU3SOXHSJGugWNZGHKJqgZ/SK3qwn68V6tz7mrStWMXOE/sj6/AGN1pVF</latexit> <latexit sha1_base64="XVa39VNO1z67Cq/lUyFwYpXmj2k=">AAAB/3icbZDNSsNAFIUn/tb6F3Xhws1gEVyVpFTUXcGNywrGFppQJtNJO3QyCTM30hK68VXcuFBx62u4822ctllo64GBj3Pv5c49YSq4Bsf5tlZW19Y3Nktb5e2d3b19++DwQSeZosyjiUhUOySaCS6ZBxwEa6eKkTgUrBUOb6b11iNTmifyHsYpC2LSlzzilICxuvbxhVvzgcdMY0PYBzaCPB1NunbFqToz4WVwC6igQs2u/eX3EprFTAIVROuO66QQ5EQBp4JNyn6mWUrokPRZx6AkZmWQzw6Y4DPj9HCUKPMk4Jn7eyInsdbjODSdMYGBXqxNzf9qnQyiqyDnMs2ASTpfFGUCQ4KnaeAeV4yCGBsgVHHzV0wHRBEKJrOyCcFdPHkZvFr1uurc1SuNepFGCZ2gU3SOXHSJGugWNZGHKJqgZ/SK3qwn68V6tz7mrStWMXOE/sj6/AGN1pVF</latexit> <latexit sha1_base64="XVa39VNO1z67Cq/lUyFwYpXmj2k=">AAAB/3icbZDNSsNAFIUn/tb6F3Xhws1gEVyVpFTUXcGNywrGFppQJtNJO3QyCTM30hK68VXcuFBx62u4822ctllo64GBj3Pv5c49YSq4Bsf5tlZW19Y3Nktb5e2d3b19++DwQSeZosyjiUhUOySaCS6ZBxwEa6eKkTgUrBUOb6b11iNTmifyHsYpC2LSlzzilICxuvbxhVvzgcdMY0PYBzaCPB1NunbFqToz4WVwC6igQs2u/eX3EprFTAIVROuO66QQ5EQBp4JNyn6mWUrokPRZx6AkZmWQzw6Y4DPj9HCUKPMk4Jn7eyInsdbjODSdMYGBXqxNzf9qnQyiqyDnMs2ASTpfFGUCQ4KnaeAeV4yCGBsgVHHzV0wHRBEKJrOyCcFdPHkZvFr1uurc1SuNepFGCZ2gU3SOXHSJGugWNZGHKJqgZ/SK3qwn68V6tz7mrStWMXOE/sj6/AGN1pVF</latexit> G Late fusion w/ single stream + LSTM CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> Cat <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> Late fusion w/ multiple streams + LSTM J Cat <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> LSTM <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> <latexit sha1_base64="9Z1R/EHi/xWz0U0JbMcWCWVq7TE=">AAAB6nicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KXjwoVGxsoQ1ls520S3c3YXcjlNC/4MWDild/kTf/jUnbg4qCDwYe780wMy+IBTfWdT+cwtLyyupacb20sbm1vVPe3bszUaIZeiwSke4E1KDgCj3LrcBOrJHKQGA7GF/kfvseteGRatlJjL6kQ8VDzqjNpavb1nW/XKlV3RnI36QCCzT75ffeIGKJRGWZoMZ0a25s/ZRqy5nAaamXGIwpG9MhdjOqqETjp7Nbp+QoUwYkjHRWypKZ+nUipdKYiQyyTkntyPz0cvE3r5vY8MxPuYoTi4rNF4WJIDYi+eNkwDUyKyYZoUzz7FbCRlRTZrN4Sv8LwTupnlfdm3qlUV+kUYQDOIRjqMEpNOASmuABgxE8wBM8O9J5dF6c13lrwVnM7MM3OG+f822Nqg==</latexit> CNN1 <latexit sha1_base64="Yg2QesPyKub94otVB9zmaGBnlrM=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0mkUL0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cftxUFHwwcDjvRlm5oWpFAZd98MprK1vbG4Vt0s7u3v7B+XDozuTZJpxnyUy0d2QGi6F4j4KlLybak7jUPJOOGnO/c4910Yk6hanKQ9iOlIiEoyilfxmqzXwBuWKV3UXIH+TCqzQHpTf+8OEZTFXyCQ1pue5KQY51SiY5LNSPzM8pWxCR7xnqaIxN0G+OHZGzqwyJFGibSkkC/XrRE5jY6ZxaDtjimPz05uLv3m9DKPLIBcqzZArtlwUZZJgQuafk6HQnKGcWkKZFvZWwsZUU4Y2n9L/QvAvqldV96ZWadRWaRThBE7hHDyoQwOuoQ0+MBDwAE/w7Cjn0XlxXpetBWc1cwzf4Lx9AmCFjeM=</latexit> <latexit sha1_base64="Yg2QesPyKub94otVB9zmaGBnlrM=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0mkUL0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cftxUFHwwcDjvRlm5oWpFAZd98MprK1vbG4Vt0s7u3v7B+XDozuTZJpxnyUy0d2QGi6F4j4KlLybak7jUPJOOGnO/c4910Yk6hanKQ9iOlIiEoyilfxmqzXwBuWKV3UXIH+TCqzQHpTf+8OEZTFXyCQ1pue5KQY51SiY5LNSPzM8pWxCR7xnqaIxN0G+OHZGzqwyJFGibSkkC/XrRE5jY6ZxaDtjimPz05uLv3m9DKPLIBcqzZArtlwUZZJgQuafk6HQnKGcWkKZFvZWwsZUU4Y2n9L/QvAvqldV96ZWadRWaRThBE7hHDyoQwOuoQ0+MBDwAE/w7Cjn0XlxXpetBWc1cwzf4Lx9AmCFjeM=</latexit> <latexit sha1_base64="Yg2QesPyKub94otVB9zmaGBnlrM=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0mkUL0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cftxUFHwwcDjvRlm5oWpFAZd98MprK1vbG4Vt0s7u3v7B+XDozuTZJpxnyUy0d2QGi6F4j4KlLybak7jUPJOOGnO/c4910Yk6hanKQ9iOlIiEoyilfxmqzXwBuWKV3UXIH+TCqzQHpTf+8OEZTFXyCQ1pue5KQY51SiY5LNSPzM8pWxCR7xnqaIxN0G+OHZGzqwyJFGibSkkC/XrRE5jY6ZxaDtjimPz05uLv3m9DKPLIBcqzZArtlwUZZJgQuafk6HQnKGcWkKZFvZWwsZUU4Y2n9L/QvAvqldV96ZWadRWaRThBE7hHDyoQwOuoQ0+MBDwAE/w7Cjn0XlxXpetBWc1cwzf4Lx9AmCFjeM=</latexit> CNN2 <latexit sha1_base64="+nS47kivtNZKKFF3UWUK4g7dndA=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0lLQb0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cZv2oKLgg4HHezPMzAsSKQy67odTWFvf2Nwqbpd2dvf2D8qHR3cmTjXjHotlrHsBNVwKxT0UKHkv0ZxGgeTdYNpa+N17ro2I1S3OEu5HdKxEKBhFK3mtdntYH5Yrtaqbg/xNKrBCZ1h+H4xilkZcIZPUmH7NTdDPqEbBJJ+XBqnhCWVTOuZ9SxWNuPGz/Ng5ObPKiISxtqWQ5OrXiYxGxsyiwHZGFCfmp7cQf/P6KYaXfiZUkiJXbLkoTCXBmCw+JyOhOUM5s4QyLeythE2opgxtPqX/heDVq1dV96ZRaTZWaRThBE7hHGpwAU24hg54wEDAAzzBs6OcR+fFeV22FpzVzDF8g/P2CWIIjeQ=</latexit> <latexit sha1_base64="+nS47kivtNZKKFF3UWUK4g7dndA=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0lLQb0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cZv2oKLgg4HHezPMzAsSKQy67odTWFvf2Nwqbpd2dvf2D8qHR3cmTjXjHotlrHsBNVwKxT0UKHkv0ZxGgeTdYNpa+N17ro2I1S3OEu5HdKxEKBhFK3mtdntYH5Yrtaqbg/xNKrBCZ1h+H4xilkZcIZPUmH7NTdDPqEbBJJ+XBqnhCWVTOuZ9SxWNuPGz/Ng5ObPKiISxtqWQ5OrXiYxGxsyiwHZGFCfmp7cQf/P6KYaXfiZUkiJXbLkoTCXBmCw+JyOhOUM5s4QyLeythE2opgxtPqX/heDVq1dV96ZRaTZWaRThBE7hHGpwAU24hg54wEDAAzzBs6OcR+fFeV22FpzVzDF8g/P2CWIIjeQ=</latexit> <latexit sha1_base64="+nS47kivtNZKKFF3UWUK4g7dndA=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0lLQb0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cZv2oKLgg4HHezPMzAsSKQy67odTWFvf2Nwqbpd2dvf2D8qHR3cmTjXjHotlrHsBNVwKxT0UKHkv0ZxGgeTdYNpa+N17ro2I1S3OEu5HdKxEKBhFK3mtdntYH5Yrtaqbg/xNKrBCZ1h+H4xilkZcIZPUmH7NTdDPqEbBJJ+XBqnhCWVTOuZ9SxWNuPGz/Ng5ObPKiISxtqWQ5OrXiYxGxsyiwHZGFCfmp7cQf/P6KYaXfiZUkiJXbLkoTCXBmCw+JyOhOUM5s4QyLeythE2opgxtPqX/heDVq1dV96ZRaTZWaRThBE7hHGpwAU24hg54wEDAAzzBs6OcR+fFeV22FpzVzDF8g/P2CWIIjeQ=</latexit> CNN3 <latexit sha1_base64="MRpyG4vMz4Af3BUF/mK5NMByDRY=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0m1oN4KvXgqFYwttKFstpN26WYTdjdCCf0NXjyoePUPefPfuGl7UFHwwcDjvRlm5gWJ4Nq47odTWFldW98obpa2tnd298r7B3c6ThVDj8UiVt2AahRcome4EdhNFNIoENgJJs3c79yj0jyWt2aaoB/RkeQhZ9RYyWu2WoPzQblSq7pzkL9JBZZoD8rv/WHM0gilYYJq3au5ifEzqgxnAmelfqoxoWxCR9izVNIItZ/Nj52RE6sMSRgrW9KQufp1IqOR1tMosJ0RNWP908vF37xeasJLP+MySQ1KtlgUpoKYmOSfkyFXyIyYWkKZ4vZWwsZUUWZsPqX/heCdVa+q7k290qgv0yjCERzDKdTgAhpwDW3wgAGHB3iCZ0c6j86L87poLTjLmUP4BuftE2OLjeU=</latexit> <latexit sha1_base64="MRpyG4vMz4Af3BUF/mK5NMByDRY=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0m1oN4KvXgqFYwttKFstpN26WYTdjdCCf0NXjyoePUPefPfuGl7UFHwwcDjvRlm5gWJ4Nq47odTWFldW98obpa2tnd298r7B3c6ThVDj8UiVt2AahRcome4EdhNFNIoENgJJs3c79yj0jyWt2aaoB/RkeQhZ9RYyWu2WoPzQblSq7pzkL9JBZZoD8rv/WHM0gilYYJq3au5ifEzqgxnAmelfqoxoWxCR9izVNIItZ/Nj52RE6sMSRgrW9KQufp1IqOR1tMosJ0RNWP908vF37xeasJLP+MySQ1KtlgUpoKYmOSfkyFXyIyYWkKZ4vZWwsZUUWZsPqX/heCdVa+q7k290qgv0yjCERzDKdTgAhpwDW3wgAGHB3iCZ0c6j86L87poLTjLmUP4BuftE2OLjeU=</latexit> <latexit sha1_base64="MRpyG4vMz4Af3BUF/mK5NMByDRY=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0m1oN4KvXgqFYwttKFstpN26WYTdjdCCf0NXjyoePUPefPfuGl7UFHwwcDjvRlm5gWJ4Nq47odTWFldW98obpa2tnd298r7B3c6ThVDj8UiVt2AahRcome4EdhNFNIoENgJJs3c79yj0jyWt2aaoB/RkeQhZ9RYyWu2WoPzQblSq7pzkL9JBZZoD8rv/WHM0gilYYJq3au5ifEzqgxnAmelfqoxoWxCR9izVNIItZ/Nj52RE6sMSRgrW9KQufp1IqOR1tMosJ0RNWP908vF37xeasJLP+MySQ1KtlgUpoKYmOSfkyFXyIyYWkKZ4vZWwsZUUWZsPqX/heCdVa+q7k290qgv0yjCERzDKdTgAhpwDW3wgAGHB3iCZ0c6j86L87poLTjLmUP4BuftE2OLjeU=</latexit> CNN4 <latexit sha1_base64="vgydTP692IoUbgnCL0zDdRFc/Rk=">AAAB63icjVDLSgNBEOzxGeMr6tHLYBA8hY0E1FsgF08hgmsCyRJmJ73JkNnZZWZWCEu+wYsHFa/+kDf/xsnjoKJgQUNR1U13V5hKYaznfZCV1bX1jc3CVnF7Z3dvv3RweGeSTHP0eSIT3QmZQSkU+lZYiZ1UI4tDie1w3Jj57XvURiTq1k5SDGI2VCISnFkn+Y1ms1/rl8rVijcH/ZuUYYlWv/TeGyQ8i1FZLpkx3aqX2iBn2goucVrsZQZTxsdsiF1HFYvRBPn82Ck9dcqARol2pSydq18nchYbM4lD1xkzOzI/vZn4m9fNbHQZ5EKlmUXFF4uiTFKb0NnndCA0cisnjjCuhbuV8hHTjFuXT/F/IfjnlauKd1Mr12vLNApwDCdwBlW4gDpcQwt84CDgAZ7gmSjySF7I66J1hSxnjuAbyNsnZQ6N5g==</latexit> <latexit sha1_base64="vgydTP692IoUbgnCL0zDdRFc/Rk=">AAAB63icjVDLSgNBEOzxGeMr6tHLYBA8hY0E1FsgF08hgmsCyRJmJ73JkNnZZWZWCEu+wYsHFa/+kDf/xsnjoKJgQUNR1U13V5hKYaznfZCV1bX1jc3CVnF7Z3dvv3RweGeSTHP0eSIT3QmZQSkU+lZYiZ1UI4tDie1w3Jj57XvURiTq1k5SDGI2VCISnFkn+Y1ms1/rl8rVijcH/ZuUYYlWv/TeGyQ8i1FZLpkx3aqX2iBn2goucVrsZQZTxsdsiF1HFYvRBPn82Ck9dcqARol2pSydq18nchYbM4lD1xkzOzI/vZn4m9fNbHQZ5EKlmUXFF4uiTFKb0NnndCA0cisnjjCuhbuV8hHTjFuXT/F/IfjnlauKd1Mr12vLNApwDCdwBlW4gDpcQwt84CDgAZ7gmSjySF7I66J1hSxnjuAbyNsnZQ6N5g==</latexit> <latexit sha1_base64="vgydTP692IoUbgnCL0zDdRFc/Rk=">AAAB63icjVDLSgNBEOzxGeMr6tHLYBA8hY0E1FsgF08hgmsCyRJmJ73JkNnZZWZWCEu+wYsHFa/+kDf/xsnjoKJgQUNR1U13V5hKYaznfZCV1bX1jc3CVnF7Z3dvv3RweGeSTHP0eSIT3QmZQSkU+lZYiZ1UI4tDie1w3Jj57XvURiTq1k5SDGI2VCISnFkn+Y1ms1/rl8rVijcH/ZuUYYlWv/TeGyQ8i1FZLpkx3aqX2iBn2goucVrsZQZTxsdsiF1HFYvRBPn82Ck9dcqARol2pSydq18nchYbM4lD1xkzOzI/vZn4m9fNbHQZ5EKlmUXFF4uiTFKb0NnndCA0cisnjjCuhbuV8hHTjFuXT/F/IfjnlauKd1Mr12vLNApwDCdwBlW4gDpcQwt84CDgAZ7gmSjySF7I66J1hSxnjuAbyNsnZQ6N5g==</latexit> F Late fusion w/ single stream CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> Cat <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> CNN <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> <latexit sha1_base64="ShW1PHfXpbXOQxWLdrslYn2rAdk=">AAAB6XicjVBNS8NAEJ3Ur1q/qh69LBbBU0mloN4KvXgqFY0ttKFstpt26WYTdidCCf0JXjyoePUfefPfuGl7UFHwwcDjvRlm5gWJFAZd98MprKyurW8UN0tb2zu7e+X9gzsTp5pxj8Uy1t2AGi6F4h4KlLybaE6jQPJOMGnmfueeayNidYvThPsRHSkRCkbRSjfNVmtQrtSq7hzkb1KBJdqD8nt/GLM04gqZpMb0am6CfkY1Cib5rNRPDU8om9AR71mqaMSNn81PnZETqwxJGGtbCslc/TqR0ciYaRTYzoji2Pz0cvE3r5dieOFnQiUpcsUWi8JUEoxJ/jcZCs0ZyqkllGlhbyVsTDVlaNMp/S8E76x6WXWv65VGfZlGEY7gGE6hBufQgCtogwcMRvAAT/DsSOfReXFeF60FZzlzCN/gvH0COtmNPw==</latexit> I Late fusion w/ multiple streams Cat <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> <latexit sha1_base64="KqbbebZMcS8vzbDadPSuhthAiF4=">AAAB6XicjVDLSgNBEOyJrxhfUY9eBoPgKWwkoN4CuXiM6JpAsoTZyWwyZHZ2mekVwpJP8OJBxat/5M2/cfI4qChY0FBUddPdFaZKWvS8D1JYWV1b3yhulra2d3b3yvsHdzbJDBc+T1RiOiGzQkktfJSoRCc1gsWhEu1w3Jz57XthrEz0LU5SEcRsqGUkOUMn3TQZ9suVWtWbg/5NKrBEq19+7w0SnsVCI1fM2m7NSzHImUHJlZiWepkVKeNjNhRdRzWLhQ3y+alTeuKUAY0S40ojnatfJ3IWWzuJQ9cZMxzZn95M/M3rZhhdBLnUaYZC88WiKFMUEzr7mw6kERzVxBHGjXS3Uj5ihnF06ZT+F4J/Vr2setf1SqO+TKMIR3AMp1CDc2jAFbTABw5DeIAneCaKPJIX8rpoLZDlzCF8A3n7BJEXjXg=</latexit> CNN1 <latexit sha1_base64="Yg2QesPyKub94otVB9zmaGBnlrM=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0mkUL0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cftxUFHwwcDjvRlm5oWpFAZd98MprK1vbG4Vt0s7u3v7B+XDozuTZJpxnyUy0d2QGi6F4j4KlLybak7jUPJOOGnO/c4910Yk6hanKQ9iOlIiEoyilfxmqzXwBuWKV3UXIH+TCqzQHpTf+8OEZTFXyCQ1pue5KQY51SiY5LNSPzM8pWxCR7xnqaIxN0G+OHZGzqwyJFGibSkkC/XrRE5jY6ZxaDtjimPz05uLv3m9DKPLIBcqzZArtlwUZZJgQuafk6HQnKGcWkKZFvZWwsZUU4Y2n9L/QvAvqldV96ZWadRWaRThBE7hHDyoQwOuoQ0+MBDwAE/w7Cjn0XlxXpetBWc1cwzf4Lx9AmCFjeM=</latexit> <latexit sha1_base64="Yg2QesPyKub94otVB9zmaGBnlrM=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0mkUL0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cftxUFHwwcDjvRlm5oWpFAZd98MprK1vbG4Vt0s7u3v7B+XDozuTZJpxnyUy0d2QGi6F4j4KlLybak7jUPJOOGnO/c4910Yk6hanKQ9iOlIiEoyilfxmqzXwBuWKV3UXIH+TCqzQHpTf+8OEZTFXyCQ1pue5KQY51SiY5LNSPzM8pWxCR7xnqaIxN0G+OHZGzqwyJFGibSkkC/XrRE5jY6ZxaDtjimPz05uLv3m9DKPLIBcqzZArtlwUZZJgQuafk6HQnKGcWkKZFvZWwsZUU4Y2n9L/QvAvqldV96ZWadRWaRThBE7hHDyoQwOuoQ0+MBDwAE/w7Cjn0XlxXpetBWc1cwzf4Lx9AmCFjeM=</latexit> <latexit sha1_base64="Yg2QesPyKub94otVB9zmaGBnlrM=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0mkUL0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cftxUFHwwcDjvRlm5oWpFAZd98MprK1vbG4Vt0s7u3v7B+XDozuTZJpxnyUy0d2QGi6F4j4KlLybak7jUPJOOGnO/c4910Yk6hanKQ9iOlIiEoyilfxmqzXwBuWKV3UXIH+TCqzQHpTf+8OEZTFXyCQ1pue5KQY51SiY5LNSPzM8pWxCR7xnqaIxN0G+OHZGzqwyJFGibSkkC/XrRE5jY6ZxaDtjimPz05uLv3m9DKPLIBcqzZArtlwUZZJgQuafk6HQnKGcWkKZFvZWwsZUU4Y2n9L/QvAvqldV96ZWadRWaRThBE7hHDyoQwOuoQ0+MBDwAE/w7Cjn0XlxXpetBWc1cwzf4Lx9AmCFjeM=</latexit> CNN2 <latexit sha1_base64="+nS47kivtNZKKFF3UWUK4g7dndA=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0lLQb0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cZv2oKLgg4HHezPMzAsSKQy67odTWFvf2Nwqbpd2dvf2D8qHR3cmTjXjHotlrHsBNVwKxT0UKHkv0ZxGgeTdYNpa+N17ro2I1S3OEu5HdKxEKBhFK3mtdntYH5Yrtaqbg/xNKrBCZ1h+H4xilkZcIZPUmH7NTdDPqEbBJJ+XBqnhCWVTOuZ9SxWNuPGz/Ng5ObPKiISxtqWQ5OrXiYxGxsyiwHZGFCfmp7cQf/P6KYaXfiZUkiJXbLkoTCXBmCw+JyOhOUM5s4QyLeythE2opgxtPqX/heDVq1dV96ZRaTZWaRThBE7hHGpwAU24hg54wEDAAzzBs6OcR+fFeV22FpzVzDF8g/P2CWIIjeQ=</latexit> <latexit sha1_base64="+nS47kivtNZKKFF3UWUK4g7dndA=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0lLQb0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cZv2oKLgg4HHezPMzAsSKQy67odTWFvf2Nwqbpd2dvf2D8qHR3cmTjXjHotlrHsBNVwKxT0UKHkv0ZxGgeTdYNpa+N17ro2I1S3OEu5HdKxEKBhFK3mtdntYH5Yrtaqbg/xNKrBCZ1h+H4xilkZcIZPUmH7NTdDPqEbBJJ+XBqnhCWVTOuZ9SxWNuPGz/Ng5ObPKiISxtqWQ5OrXiYxGxsyiwHZGFCfmp7cQf/P6KYaXfiZUkiJXbLkoTCXBmCw+JyOhOUM5s4QyLeythE2opgxtPqX/heDVq1dV96ZRaTZWaRThBE7hHGpwAU24hg54wEDAAzzBs6OcR+fFeV22FpzVzDF8g/P2CWIIjeQ=</latexit> <latexit sha1_base64="+nS47kivtNZKKFF3UWUK4g7dndA=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0lLQb0VevFUKhhbaEPZbDft0s0m7E6EEvobvHhQ8eof8ua/cZv2oKLgg4HHezPMzAsSKQy67odTWFvf2Nwqbpd2dvf2D8qHR3cmTjXjHotlrHsBNVwKxT0UKHkv0ZxGgeTdYNpa+N17ro2I1S3OEu5HdKxEKBhFK3mtdntYH5Yrtaqbg/xNKrBCZ1h+H4xilkZcIZPUmH7NTdDPqEbBJJ+XBqnhCWVTOuZ9SxWNuPGz/Ng5ObPKiISxtqWQ5OrXiYxGxsyiwHZGFCfmp7cQf/P6KYaXfiZUkiJXbLkoTCXBmCw+JyOhOUM5s4QyLeythE2opgxtPqX/heDVq1dV96ZRaTZWaRThBE7hHGpwAU24hg54wEDAAzzBs6OcR+fFeV22FpzVzDF8g/P2CWIIjeQ=</latexit> CNN3 <latexit sha1_base64="MRpyG4vMz4Af3BUF/mK5NMByDRY=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0m1oN4KvXgqFYwttKFstpN26WYTdjdCCf0NXjyoePUPefPfuGl7UFHwwcDjvRlm5gWJ4Nq47odTWFldW98obpa2tnd298r7B3c6ThVDj8UiVt2AahRcome4EdhNFNIoENgJJs3c79yj0jyWt2aaoB/RkeQhZ9RYyWu2WoPzQblSq7pzkL9JBZZoD8rv/WHM0gilYYJq3au5ifEzqgxnAmelfqoxoWxCR9izVNIItZ/Nj52RE6sMSRgrW9KQufp1IqOR1tMosJ0RNWP908vF37xeasJLP+MySQ1KtlgUpoKYmOSfkyFXyIyYWkKZ4vZWwsZUUWZsPqX/heCdVa+q7k290qgv0yjCERzDKdTgAhpwDW3wgAGHB3iCZ0c6j86L87poLTjLmUP4BuftE2OLjeU=</latexit> <latexit sha1_base64="MRpyG4vMz4Af3BUF/mK5NMByDRY=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0m1oN4KvXgqFYwttKFstpN26WYTdjdCCf0NXjyoePUPefPfuGl7UFHwwcDjvRlm5gWJ4Nq47odTWFldW98obpa2tnd298r7B3c6ThVDj8UiVt2AahRcome4EdhNFNIoENgJJs3c79yj0jyWt2aaoB/RkeQhZ9RYyWu2WoPzQblSq7pzkL9JBZZoD8rv/WHM0gilYYJq3au5ifEzqgxnAmelfqoxoWxCR9izVNIItZ/Nj52RE6sMSRgrW9KQufp1IqOR1tMosJ0RNWP908vF37xeasJLP+MySQ1KtlgUpoKYmOSfkyFXyIyYWkKZ4vZWwsZUUWZsPqX/heCdVa+q7k290qgv0yjCERzDKdTgAhpwDW3wgAGHB3iCZ0c6j86L87poLTjLmUP4BuftE2OLjeU=</latexit> <latexit sha1_base64="MRpyG4vMz4Af3BUF/mK5NMByDRY=">AAAB63icjVBNS8NAEJ3Ur1q/qh69LBbBU0m1oN4KvXgqFYwttKFstpN26WYTdjdCCf0NXjyoePUPefPfuGl7UFHwwcDjvRlm5gWJ4Nq47odTWFldW98obpa2tnd298r7B3c6ThVDj8UiVt2AahRcome4EdhNFNIoENgJJs3c79yj0jyWt2aaoB/RkeQhZ9RYyWu2WoPzQblSq7pzkL9JBZZoD8rv/WHM0gilYYJq3au5ifEzqgxnAmelfqoxoWxCR9izVNIItZ/Nj52RE6sMSRgrW9KQufp1IqOR1tMosJ0RNWP908vF37xeasJLP+MySQ1KtlgUpoKYmOSfkyFXyIyYWkKZ4vZWwsZUUWZsPqX/heCdVa+q7k290qgv0yjCERzDKdTgAhpwDW3wgAGHB3iCZ0c6j86L87poLTjLmUP4BuftE2OLjeU=</latexit> CNN4 <latexit sha1_base64="vgydTP692IoUbgnCL0zDdRFc/Rk=">AAAB63icjVDLSgNBEOzxGeMr6tHLYBA8hY0E1FsgF08hgmsCyRJmJ73JkNnZZWZWCEu+wYsHFa/+kDf/xsnjoKJgQUNR1U13V5hKYaznfZCV1bX1jc3CVnF7Z3dvv3RweGeSTHP0eSIT3QmZQSkU+lZYiZ1UI4tDie1w3Jj57XvURiTq1k5SDGI2VCISnFkn+Y1ms1/rl8rVijcH/ZuUYYlWv/TeGyQ8i1FZLpkx3aqX2iBn2goucVrsZQZTxsdsiF1HFYvRBPn82Ck9dcqARol2pSydq18nchYbM4lD1xkzOzI/vZn4m9fNbHQZ5EKlmUXFF4uiTFKb0NnndCA0cisnjjCuhbuV8hHTjFuXT/F/IfjnlauKd1Mr12vLNApwDCdwBlW4gDpcQwt84CDgAZ7gmSjySF7I66J1hSxnjuAbyNsnZQ6N5g==</latexit> <latexit sha1_base64="vgydTP692IoUbgnCL0zDdRFc/Rk=">AAAB63icjVDLSgNBEOzxGeMr6tHLYBA8hY0E1FsgF08hgmsCyRJmJ73JkNnZZWZWCEu+wYsHFa/+kDf/xsnjoKJgQUNR1U13V5hKYaznfZCV1bX1jc3CVnF7Z3dvv3RweGeSTHP0eSIT3QmZQSkU+lZYiZ1UI4tDie1w3Jj57XvURiTq1k5SDGI2VCISnFkn+Y1ms1/rl8rVijcH/ZuUYYlWv/TeGyQ8i1FZLpkx3aqX2iBn2goucVrsZQZTxsdsiF1HFYvRBPn82Ck9dcqARol2pSydq18nchYbM4lD1xkzOzI/vZn4m9fNbHQZ5EKlmUXFF4uiTFKb0NnndCA0cisnjjCuhbuV8hHTjFuXT/F/IfjnlauKd1Mr12vLNApwDCdwBlW4gDpcQwt84CDgAZ7gmSjySF7I66J1hSxnjuAbyNsnZQ6N5g==</latexit> <latexit sha1_base64="vgydTP692IoUbgnCL0zDdRFc/Rk=">AAAB63icjVDLSgNBEOzxGeMr6tHLYBA8hY0E1FsgF08hgmsCyRJmJ73JkNnZZWZWCEu+wYsHFa/+kDf/xsnjoKJgQUNR1U13V5hKYaznfZCV1bX1jc3CVnF7Z3dvv3RweGeSTHP0eSIT3QmZQSkU+lZYiZ1UI4tDie1w3Jj57XvURiTq1k5SDGI2VCISnFkn+Y1ms1/rl8rVijcH/ZuUYYlWv/TeGyQ8i1FZLpkx3aqX2iBn2goucVrsZQZTxsdsiF1HFYvRBPn82Ck9dcqARol2pSydq18nchYbM4lD1xkzOzI/vZn4m9fNbHQZ5EKlmUXFF4uiTFKb0NnndCA0cisnjjCuhbuV8hHTjFuXT/F/IfjnlauKd1Mr12vLNApwDCdwBlW4gDpcQwt84CDgAZ7gmSjySF7I66J1hSxnjuAbyNsnZQ6N5g==</latexit> P <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> <latexit sha1_base64="i3Va4qcItgjb0qXbrfXCNvVAt44=">AAAB53icjVDLSgNBEOyNrxhfUY9eBoPgKWxEUG8BLx4TcE0gWWR20puMmZ1dZnqFEPIFXjyoePWXvPk3Th4HFQULBoqqarqnokxJS77/4RWWlldW14rrpY3Nre2d8u7ejU1zIzAQqUpNO+IWldQYkCSF7cwgTyKFrWh4OfVb92isTPU1jTIME97XMpaCk5OajdtypVb1Z2B/kwos4PLv3V4q8gQ1CcWt7dT8jMIxNySFwkmpm1vMuBjyPnYc1TxBG45nh07YkVN6LE6Ne5rYTP06MeaJtaMkcsmE08D+9Kbib14np/g8HEud5YRazBfFuWKUsumvWU8aFKRGjnBhpLuViQE3XJDrpvS/EoKT6kXVb55W6qeLNopwAIdwDDU4gzpcQQMCEIDwAE/w7N15j96L9zqPFrzFzD58g/f2CRgBjJw=</latexit> Fig. 2. Used architectures. Model complexity and run time are specified in Table 1. location. Each patch is processed on a separated branch of a CNN, yielding multiple-scale features which are then combined for the final prediction. Instead of extracting multiple patches at different scales, Kong et al. [12] use a CNN with a 2-dimensional long-short term (LSTM) architecture [4] to learn spatial dependencies of image patches and their neighbours. Incorporating multi-scale and contextual information into a patch-wise classification scheme is still an open problem. A systematic comparison of different network architectures is necessary to establish how visual context should be utilised in whole slide image segmentation. 3 Methods Comparative methods. The 10 different architectures that are being used in this study are presented in Figure 2. These can be categorised into three groups: 1) those that operates at a particular image resolution (A, B, C, D, Improving Whole Slide Segmentation Through Visual Context 5 Table 1. Classification accuracy as measured by the F1-measure. Bold indi- cates the best performance. Green, blue, yellow, and red colour codings indicate that the results are within 97.5%, 95%, 90%, and 85% of the best performance, respectively. This colour coding scheme can be used to rank the methods (bold = 1, green = 2, blue = 3, yellow = 4, red = 5, and no colour = 6). The overall ranking is summarised by the rank-sum. A total running time is measured on the test set of the prostate cancer data. Dataset Class Method A B C D E F G H I J Prostate Lumen 0.728 0.663 0.705 0.716 0.739 0.738 0.748 0.713 0.722 0.758 Stroma 0.797 0.855 0.849 0.790 0.875 0.869 0.884 0.891 0.862 0.883 Benign 0.508 0.646 0.712 0.717 0.734 0.745 0.766 0.763 0.765 0.782 Tumour 0.562 0.653 0.629 0.579 0.699 0.687 0.728 0.746 0.674 0.712 Breast Normal 0.501 0.468 0.523 0.513 0.509 0.603 0.573 0.252 0.241 0.323 Benign 0.453 0.468 0.482 0.444 0.410 0.369 0.423 0.489 0.333 0.437 InSitu 0.468 0.476 0.486 0.533 0.615 0.614 0.581 0.286 0.311 0.452 Invasive 0.401 0.477 0.430 0.540 0.557 0.548 0.576 0.520 0.446 0.580 Rank-sum (Prostate) 20 19 17 19 13 13 8 8 12 7 Rank-sum (Breast) 22 21 19 18 16 13 14 18 24 18 Total rank-sum 42 40 36 37 29 26 22 26 36 25 No. of parameters 7.2M 7.2M 7.2M 7.2M 7.3M 8.0M 10.1M 19.8M 28.9M 31.0M Running time (s) 7.16 7.16 7.16 7.16 7.21 7.61 7.62 35.59 7.60 7.70 While model H yields the top performance for selected classes, it also per- forms rather poorly on others. Given that this model performs extremely well on detecting stroma in prostate tissue, one could argue that it specialises on capturing certain texture patters extremly well. When comparing models G and J we can make some interesting observations. On the given data sets model G performs consistently well in all of the tissue classes and has the lowest accumu- lative rank. Only considering the prostate samples model J is clearly the best. However, the performance of this model degrades on the breast cancer cases. Here, the interplay between model complexity and size of the data set needs to be taken into account. Later we discuss this issue in more detail. Overall, these results support our hypothesis that visual context and scale matters in histology ⼊⼒画像のサイズや倍率,ネットワーク(CNN, LSTM)の組み合わせを 変えて10通りの学習を検討 to learn the global interdependence of various structures in dif- ferent lesion categories. The performance of our system is evalu- ated on a large breast histopathology cohort comprising 221 WSIs from 122 patients. 2 Methods 2.1 Overview of the System The main challenge in the design of our classification frame- work is that the appearance of many benign diseases of the breast (e.g., usual ductal hyperplasia) mimics that of DCIS, hence requiring accurate texture analysis at the cellular level. Such analysis, however, is not sufficient for discrimination of DCIS from IDC. DCIS and IDC may appear identical on cellular examination but are different in their growth patterns, which can only be captured through the inclusion of larger image patches containing more information about the global tissue architec- ture. Because of computational constraints, however, it is not feasible to train a deep CNN with large patches at high resolu- tion that contain enough context. Our method for classification of breast histopathology WSIs overcomes these problems through sequential analysis with a stack of CNNs. The key components of our classification frame- work, including the CNN used for classification of high-resolu- tion patches, the stacked CNN for producing dense prediction maps, and a WSI labeling module, are detailed in the following sections. 2.2 Deep Convolutional Neural Network for Classification of Small High-Resolution Patches Inspired by the recent successes of deep residual networks26 for image classification, we trained and evaluated the performance of this CNN for classification of small high-resolution patches into normal/benign, DCIS, and IDC. We applied an adaptation of the ResNet architecture called wide ResNet, as proposed by Zagoruyko and Komodakis.27 This architecture has two hyper- parameters: N and K determining the depth and width of the network, respectively. We empirically chose N ¼ 4 and K ¼ 2 as a tradeoff between model capacity, training speed, and memory usage. Hereafter, we denote this network as WRN- 4-2 (see Fig. 2). This network takes as input patches of size 224 × 224. Zero padding was used before each convolutional layer to keep the spatial dimension of feature maps constant after convolution. The goal of this step was to transfer the highly informative feature representations learned by this network produced at its last convolutional layer to a stacked network, which is described next. 2.3 Context-Aware Stacked Convolutional Neural Network In order to increase the context available for dense prediction, we stack a second CNN on top of the last convolutional layer of the previously trained WRN-4-2 network. The architecture of the stacked network, as shown in Fig. 2, is a hybrid between Input conv 3@16, /2 RB 3@32, /2 RB 3@64, /2 RB 3@128, /2 RB 3@128 RB 3@128 RB 3@128 global average pooling softmax RB 3@32 RB 3@32 RB 3@32 RB 3@64 RB 3@64 RB 3@64 Input conv 3@16, /2 RB 3@32, /2 RB 3@64, /2 RB 3@128, /2 RB 3@128 RB 3@128 RB 3@128 softmax RB 3@32 RB 3@32 RB 3@32 RB 3@64 RB 3@64 RB 3@64 batch norm ReLU conv 3 batch norm ReLU conv 3 conv 1 conv 3@128 RB 3@128 RB 3@128 RB 3@128 RB 3@128 conv 3@192 conv 3@192 max pooling 2 max pooling 2 conv 12@256 conv 1@3 max pooling 2 (a) (b) (c) Residual convolution group (N = 4) Bejnordi et al.: Context-aware stacked convolutional neural networks for classification of breast. . . to learn the global interdependence of various structures in dif- ferent lesion categories. The performance of our system is evalu- ated on a large breast histopathology cohort comprising 221 WSIs from 122 patients. 2 Methods 2.1 Overview of the System The main challenge in the design of our classification frame- work is that the appearance of many benign diseases of the breast (e.g., usual ductal hyperplasia) mimics that of DCIS, hence requiring accurate texture analysis at the cellular level. Such analysis, however, is not sufficient for discrimination of DCIS from IDC. DCIS and IDC may appear identical on cellular examination but are different in their growth patterns, which can only be captured through the inclusion of larger image patches containing more information about the global tissue architec- ture. Because of computational constraints, however, it is not feasible to train a deep CNN with large patches at high resolu- tion that contain enough context. Our method for classification of breast histopathology WSIs overcomes these problems through sequential analysis with a stack of CNNs. The key components of our classification frame- work, including the CNN used for classification of high-resolu- tion patches, the stacked CNN for producing dense prediction maps, and a WSI labeling module, are detailed in the following sections. 2.2 Deep Convolutional Neural Network for Classification of Small High-Resolution Patches Inspired by the recent successes of deep residual networks26 for image classification, we trained and evaluated the performance of this CNN for classification of small high-resolution patches into normal/benign, DCIS, and IDC. We applied an adaptation of the ResNet architecture called wide ResNet, as proposed by Zagoruyko and Komodakis.27 This architecture has two hyper- parameters: N and K determining the depth and width of the network, respectively. We empirically chose N ¼ 4 and K ¼ 2 as a tradeoff between model capacity, training speed, and memory usage. Hereafter, we denote this network as WRN- 4-2 (see Fig. 2). This network takes as input patches of size 224 × 224. Zero padding was used before each convolutional layer to keep the spatial dimension of feature maps constant after convolution. The goal of this step was to transfer the highly informative feature representations learned by this network produced at its last convolutional layer to a stacked network, which is described next. 2.3 Context-Aware Stacked Convolutional Neural Network In order to increase the context available for dense prediction, we stack a second CNN on top of the last convolutional layer of the previously trained WRN-4-2 network. The architecture of the stacked network, as shown in Fig. 2, is a hybrid between Input conv 3@16, /2 RB 3@32, /2 RB 3@64, /2 RB 3@128, /2 RB 3@128 RB 3@128 RB 3@128 global average pooling softmax RB 3@32 RB 3@32 RB 3@32 RB 3@64 RB 3@64 RB 3@64 Input conv 3@16, /2 RB 3@32, /2 RB 3@64, /2 RB 3@128, /2 RB 3@128 RB 3@128 RB 3@128 softmax RB 3@32 RB 3@32 RB 3@32 RB 3@64 RB 3@64 RB 3@64 batch norm ReLU conv 3 batch norm ReLU conv 3 conv 1 conv 3@128 RB 3@128 RB 3@128 RB 3@128 RB 3@128 conv 3@192 conv 3@192 max pooling 2 max pooling 2 conv 12@256 conv 1@3 max pooling 2 (a) (b) (c) Residual convolution group (N = 4) Bejnordi et al.: Context-aware stacked convolutional neural networks for classification of breast. . . Bejnordi et al. J.Med. Imag. 2017 Context-aware stacked CNN 224 x 224 768 x 768 ⼩さいパッチの学習に⽤いたネットワークと重みを 使って⼤きいパッチの学習を⾏う ⾃施設の乳癌セットで分類タスクのaccuracy 3クラス 0.9135, 2クラス 0.962 Improving Whole Slide SegmentaIon Through Visual Context - A SystemaIc Study Context-aware stacked convoluIonal neural networks for classificaIon of breast carcinomas in whole-slide histopathology images 64 x 64 512 x 512
  4. Figure 4. Box plots of (a) Ratio of the frequency

    of first two bins of BAM to the overall frequency, (b) Average BAM and (c) BAM entropy. The box plots are generated after feature postprocessing. Our classification into three distinct populations (normal, low grade and high grade) has p-value less than 0.001 for each of the three criteria (Regularity Index, average BAM value and BAM entropy). Figure 5. An overall flowchart of our methodology. Figure 6. Network architecture. Each slice represents a multi-channel feature map and the depth of each feature map is mentioned on the top of layers. The orange slice is the copied feature map cropped from the 2 Fig. 1. Three visual field regions of colorectal tissue which highlight the importance of larger context for correct grading. Each cell of the overlaid grid shows the 224 ⇥ 224 pixel context captured by a standard patch classifier at 20⇥ magnification. 1,792 x 1,792 pixel 224x224 pixel 4,548 x 7,520 pixel (0.275micro/pixel), 139 images normal, low grade, high grade の3段階分類のラベル Segmentation maskあり Colorectal Cancer Dataset @Warwick 使⽤したデータセット Bioimaging 2015 breast histology classificaIon challenge R.Awan et al, Scientific Rep. 2017 6 TABLE II NUMBER OF PATCHES IN EACH CLASS AND FOLD OF TRAINING DATASET. Patches of size 224⇥224 Folds Background Normal Low Grade High Grade 1 25,000 25,000 25,000 25,000 2 25,000 25,000 25,000 25,000 3 25,000 25,000 25,000 25,000 Total 75,000 75,000 75,000 75,000 Patches of size 1,792⇥1,792 1 1,750 3,500 3,500 3,500 2 1,750 3,500 3,500 3,500 3 1,750 3,500 3,500 3,500 Total 5,250 10,500 10,500 10,500 of an image weight. Third, multi-task learning based training with the help of an auxiliary block by using joint classification and segmentation loss, Ljoint . Last, training using the same joint loss but with attention-based feature-cube to amplify the contribution of more important features in the feature- cube. The network configuration of this strategy is represented by both solid and dotted lines blocks in Fig 2. We termed these strategies as standard, weighted, auxiliary, and attention, respectively. IV. DATASETS & PERFORMANCE MEASURES In this section, we explain the dataset details used for metric is used to summarize the accuracy of different models trained using a specific setting in order to compare models trained with different context-blocks and LR-CNNs. Different colors are used to represent different rank for better illustrative visualization as shown in Table IV, V and VIII. The orange color indicates the best performing method whereas green, blue, yellow, and red colours indicate that the results are within 97.5%, 95%, 90%, and 85% of the best performing method, respectively. The rank for these colors are: orange = 1, green = 2, blue = 3, yellow = 4, red = 5, and no colour = 6. The lowest rank-sum shows the best performance. V. EXPERIMENTS & RESULTS The proposed framework is extensively evaluated and com- pared with existing approaches in three different categories, i.e. traditional patch based classifiers, existing context-aware approaches and domain oriented methods for CRC grading. The details of experimental evaluation are given in following subsections. A. Experimental Setup The CRC images are divided into patches of size 1, 792 ⇥ 1, 792, and the label of each patch is predicted using the proposed framework with a stride of 224 ⇥ 224. To avoid http://www.bioimaging2015.ineb.up.pt/challenge_overview.html 2,048 x 1,536 pixel (0.42micro/pixel), 140 images normal, benign, DCIS, invasive の4クラス分類のラベル Breast cancer diagnosis usually consists in an initial detection via palpation and regular check-ups using mammography or ultrasound imaging. The diagnosis is then followed by breast tissue biopsy if the check-up exam indicates the possibility of malignant tissue growth [4]. Breast tissue biopsies allow the pathologists to histologically assess the microscopic struc- ture and elements of the tissue. The histology allows to distinguish between normal tissue, non-malignant (benign) and malignant lesions and to perform a prognostic evaluation [5]. Benign lesions represent changes in normal structures of breast parenchyma that are not directly related with progression to malignancy. Carcinomas can be classified as in situ or inva- sive. In in situ carcinoma the cells are restrained inside the mammary ductal-lobular system, whereas in invasive carcinoma the cells spread beyond that structure. The tissue collected during the biopsy is commonly stained with hematoxylin and eosin (H&E) prior to the visual analysis performed by the specialists. During this procedure, relevant regions of whole-slide tissue scans are assessed [6]. Fig 1 shows an example of patches from whole slide images stained with H&E for each of the classes mentioned. The staining enhances nuclei (purple) and cytoplasm (pinkish), as well as other structures of interest [7]. During the analysis of the stained tissue, pathologists analyze overall tissue architecture, along with nuclei organization, density and variability. For instance, tissues with invasive car- cinoma show a distortion of the architecture as well as higher nuclei density and variability (Fig 1-D), whereas in normal tissue the architecture is maintained and the nuclei are well orga- nized (Fig 1-A). The diagnosis process using H&E stained biopsies is not trivial, and the average diagnostic concordance between specialists is approximately 75% [8]. The manual examination of histol- ogy images requires intense workload of highly specialized pathologists. The subjectivity of the application of morphological criteria in usual classification motivates the use of computer- aided diagnosis (CAD) systems to improve the diagnosis efficiency and increase the level of inter-observer agreement [9]. Related work CAD systems are embed Image Analysis and Machine Learning Methodologies developed to help physicians during the diagnosis procedure. Being a second opinion system, CAD systems reduce the workload of specialists, contributing to both diagnosis efficiency and cost reduc- tion. For that purpose, there’s often an attempt to replicate the physicians’ method. For instance, the analysis of nuclei morphology may be sufficient to classify a tissue as benign or malignant [10]. Fig 1. Examples of microscopy image patches from the used dataset [3]. Nuclei and cytoplasm appear purple and pinkish, respectively, due to the hematoxylin and eosin staining. A normal tissue; B benign abnormality; C malignant carcinoma in situ; D malignant invasive carcinoma. https://doi.org/10.1371/journal.pone.0177544.g001 Classification of breast cancer histology images using Convolutional Neural Networks PLOS ONE | https://doi.org/10.1371/journal.pone.0177544 June 1, 2017 2 / 14 Development Fund (ERDF). Teresa Arau ´jo is funded by the grant contract SFRH/BD/122365/ 2016 (Fundac ¸ão para a Ciência e a Tecnologia). Guilherme Aresta is funded by the grant contract SFRH/BD/120435/2016 (Fundac ¸ão para a Ciência e a Tecnologia). Jose ´ Rouco is funded by the grant contract SFRH/BPD/79154/2011 (Fundac ¸ão para a Ciência e a Tecnologia). Competing interests: The authors have declared that no competing interests exist. 評価基準 1. average accuracy 2. F1 score 3. rank-sum 29. Teresa Arau ́jo et al, PLOS ONE 2017 ⽂献29と同じ条件でデータセットを作成 512 x 512 pixel, 50% overlapで切り出し 2つのサイズで切り出して 新たなデータセットを作成 backgroundラベルを追加
  5. 7 C ce or d, ee TABLE V ROBUSTNESS ANALYSIS

    OF FEATURE EXTRACTORS ACROSS DIFFERENT METHODS. Methods ResNet50 (%) MobileNet(%) InceptionV3(%) Xception(%) RA-CNN 1 (Avg) 94.25±2.70 93.52±3.55 94.23±3.71 94.96±2.72 RA-CNN 1 (Max) 93.52±1.87 93.51±3.10 94.23±2.07 93.54±3.03 RA-CNN 2 (Avg) 92.08±2.08 93.52±1.78 94.96±2.72 94.96±2.72 RA-CNN 2 (Max) 95.68±3.55 93.52±3.55 92.80±2.72 93.54±3.03 RA-CNN 3 (Avg) 93.51±3.10 94.25±2.70 95.68±1.78 95.68±3.55 RA-CNN 3 (Max) 94.23±2.07 92.82±2.01 94.25±2.70 94.96±2.72 Rank-sum 12 12 10 8 C. RA-CNN based Context-Aware Learning We experimented with three context-blocks, B1 , B2 , and B3 , to train three different variation of RA-CNN, which we termed as RA-CNN 1, RA-CNN 2, and RA-CNN 3. These three RA-CNN classifiers are trained separately with all four LR- CNNs as explained in section III-F, hence giving 12 different combinations of the context-aware network. The rank-sum method is used to compare the CRC grading performance of these networks with each other and also with the LR-CNNs. The results in table IV, shows that context-aware networks achieve superior performance as compare to standard patch based classifiers (LR-CNNs). The RA-CNN 3 achieve the best Rank-sum (lowest) which shows its robustness across different representation learning networks. Other two context-aware networks also show comparable performance by remaining in the 97.5% of the best performer. TABLE IV ACCURACY COMPARISON OF THREE DIFFERENT CONTEXT-AWARE NETWORKS WITH STANDARD PATCH CLASSIFIERS. LR-CNN (Avg) Baseline RA-CNN 1 RA-CNN 2 RA-CNN 3 ResNet50 92.08±2.08 94.25±2.70 92.08±2.08 93.51±3.10 MobileNet 92.78±2.74 93.52±3.55 93.52±1.78 94.25±2.70 InceptionV3 91.37±3.55 94.23±3.71 94.96±2.72 95.68±1.78 Xception 92.09±0.98 94.96±2.72 94.96±2.72 95.68±3.55 Rank-sum 10 7 8 5 M E. T W strate explo of ea VI s Xcep the a featu colum for m achie How pooli impo basel illust CNN comb RA-C are il ison. cons for r mate 3 Context Block Context Block Context Block Convolution Softmax Context Block Features Probabilities Attention Block Segmentation Average Pool Fully Connected Auxiliary Block Classification Attention Network Multiplication Auxiliary Network 0 LR-CNN 0 Representation Aggregation Input Image Patches Features Feature Extraction Local Representation Feature Cube Fig. 2. Flow diagram of the proposed context-aware framework for CRC grading. The top row shows the local representation learning. The bottom row illustrates the network architecture for representation aggregation learning which consists of multiple context blocks and other standard layers. Dashed lines represent the blocks of a specific network design whereas solid lines represent the common blocks (see Table I for notations). the patches extracted from a high-resolution image. Due to the nature of the final classifier, this work is only capable of capturing a limited context. Bejnordi et al. [28] proposed a similar approach for breast tissue classification. They trained their network in two steps. In the first step, they used a small patch size and in the second step, they fixed the weights of half of the network to feed a larger patch for training the remaining half of the network. Their network also suffers from a limited context problem as they managed to train a network with the largest patch size of 1, 024 ⇥ 1, 024 pixels with small batch features based on the glandular morphology for prostate cancer grading. Awan et al. [11] presented a method for two-tier CRC grading based on the extent of deviation of the gland from its normal shape (circular/elliptical). They proposed a novel Best Alignment Metric (BAM) for this purpose. As a pre-processing step, CNN based gland segmentation was performed, followed by the calculation of BAM for each gland. For every image, average BAM was considered as a feature along with two more features inspired by BAM values. In the end, an SVM folds with the lowest standard deviation (Std.). TABLE III ACCURACY COMPARISON OF FOUR PATCH CLASSIFIERS. Network Fold-1 Fold-2 Fold-3 Mean Std. ResNet50 93.48 93.62 89.13 92.08 2.08 MobileNet 93.48 95.74 89.13 92.78 2.74 Inception-v3 95.65 91.49 86.96 91.37 3.55 Xception 93.48 91.49 91.30 92.09 0.98 C. RA-CNN based Context-Aware Learning We experimented with three context-blocks, B1 , B2 , and B3 , to train three different variation of RA-CNN, which we termed as RA-CNN 1, RA-CNN 2, and RA-CNN 3. These three RA-CNN classifiers are trained separately with all four LR- CNNs as explained in section III-F, hence giving 12 different combinations of the context-aware network. The rank-sum method is used to compare the CRC grading performance of these networks with each other and also with the LR-CNNs. The results in table IV, shows that context-aware networks achieve superior performance as compare to standard patch based classifiers (LR-CNNs). The RA-CNN 3 achieve the best Rank-sum (lowest) which shows its robustness across different representation learning networks. Other two context-aware networks also show comparable performance by remaining in the 97.5% of the best performer. TABLE IV ACCURACY COMPARISON OF THREE DIFFERENT CONTEXT-AWARE NETWORKS WITH STANDARD PATCH CLASSIFIERS. R R R R R CO E. W stra exp of VI Xc the fea col for ach Ho poo im bas illu CN com context-block B1 B2 B3 Local Representation network (LR-CNN) RepresentaIon AggregaIon network (RA-CNN) resnet bottleneck incepeon module ネットワークは何でも使えてflexible Feature Cubeのdimensionはネットワークによって異なる 次のRA-CNN の⼊⼒ context-block RA-CNN1 RA-CNN2 RA-CNN3 Xception based LR-CNNを以降の学習に対するbaselineとした
  6. 4パターンのtraining strategies 2. Weighted sample-based weighted loss を計算 Backgroundに対してROIが少ないパッチに重みをつける 4.

    Attention 3に1x1のconv + softmaxを挟んだAttention Blockをさらに追加 1. Standard ベースライン.Lclsを⼩さくするだけ F. Representation Aggregation for Context Learning The local representation tissue has been learned by the LR- CNN. Therefore, the task of spatial context learning from feature-cube is relatively less challenging as compared to context learning from the raw image. A cascaded set of three context blocks (C(·)) of the same type (B1 ,B2 , or B3 ) is used in RA-CNN. These context blocks are explained in section III-E. The output of C(·) is followed by a global average pooling, a fully connected, and a softmax layer to make the final prediction in the required number of classes. The final prediction Y0 from the features of input images X is computed as: Y0 = C(F0, ✓C) ! Lg p (•) ! Lf (•, ✓f0 ) ! Ls(•), (5) where ✓C and ✓f0 represent the parameters of all context blocks and the fully connected layer in RA-CNN, respectively. The proposed framework is trained end-to-end with categorical cross-entropy loss based cost function Lcls(·) which is defined as: Lcls(Y, Y0) = 1 K K X k=1 C X c=1 Y k c log2 (Y 0k c ), (6) where Y k c and Y 0k c are the ground truth and predicted proba- bilities of kth image for cth class. where ↵ is a hyper-parameter which defines the contribution of both loss functions in the final loss. Similar to patch clas- sifier, the loss function (Ljoint ) is minimized with RMSprop optimizer [38]. H. Training Strategies We trained the proposed framework in four different ways with varying ability to capture the spatial context. First, the proposed framework is trained without attention block and by minimizing the Lcls(·) loss only. This configuration is represented by solid line blocks in Fig 2. Second, the same configuration as first but trained with a sample-based weighted loss function, Lwgt(·), which give more weight to the image patches with relatively less region of interest (glandular region) as compared to the background. The weight of an image and Lwgt(·) are defined as follow, Wk = ( 1 Rk roi , if Rk roi > ↵ 1 ↵ , otherwise (10) Lwgt(Y, Y0) = 1 K K X k=1 C X c=1 WkY k c log2 (Y 0k c ), (11) where Rk roi and Wk represent the ratio of the region of interest and weight of the kth image. The ↵ is the ratio threshold, selected empirically as 0.10, sets the upper limit 3. Auxiliary Auxiliary Blockを追加して分類とセグメンテーションの両⽅のロスを計算 where L1⇥1 c and L3⇥3 c denote the convolution layers with 1⇥1 and 3 ⇥ 3 filter sizes; ✓B1 2 , ✓B2 2 , and ✓B3 2 are the parameters of different convolution layers and ✓B2 represents parameter of the whole context block for brevity. The operator represents the concatenation of feature-maps. Unlike the previous two context blocks, our third CB processes the input feature-maps in parallel with different filter sizes to capture context from varying receptive fields. Similar to the blocks in [35], it consists of multiple 1 ⇥ 1 and 3 ⇥ 3 convolution layers each followed by batch normalization and ReLU activation. A 3 ⇥ 3 average pooling layer L3⇥3 p is also used to average the local context information. The CB, B3 , is defined as: B3(F0, ✓B3 ) = [L1⇥1 c (F0, ✓B1 3 ) ! L3⇥3 c (•, ✓B2 3 ) ! L3⇥3 c (•, ✓B3 3 )] [L1⇥1 c (F0, ✓B4 3 )] [L1⇥1 c (F0, ✓B5 3 ) ! L3⇥3 c (•, ✓B6 3 )] [L3⇥3 p (F0) ! L1⇥1 c (•, ✓B7 3 )], (4) where ✓B1 3 to ✓B7 3 are the parameters of different convolution layers and ✓B3 represents parameter of the whole context block for the sake of notational simplicity. F. Representation Aggregation for Context Learning The local representation tissue has been learned by the LR- CNN. Therefore, the task of spatial context learning from feature-cube is relatively less challenging as compared to context learning from the raw image. A cascaded set of three context blocks (C(·)) of the same type (B1 ,B2 , or B3 ) is used in RA-CNN. These context blocks are explained in section III-E. The output of C(·) is followed by a global average pooling, a fully connected, and a softmax layer to make the final prediction in the required number of classes. The final prediction Y0 from the features of input images X is computed primitive structures in the input image. This will improve the convergence of proposed networks and also output the coarse patch based segmentation mask (S0 s ) along with image label (Y 0). The segmentation masks (S0) of input images X from their features F0 is defined as: S0 = C(F0, ✓C) ! L1⇥1 c (•, ✓c0 ) ! Ls(•), (7) where L1⇥1 c is a convolution layer with ✓c0 parameters. The addition of auxiliary block enables the proposes framework to learn in a multi-task setting, where the coarse segmentation- map guides the network to improve the individual patch based feature classification in addition to the prediction of the input image. This leads to a network with improved classifica- tion performance since it is minimizing both segmentation and classification loss simultaneously. The segmentation-map based loss function (Lseg ) and joint loss function (Ljoint ) are defined as: Lseg(S, S0) = 1 K K X k=1 C X c=1 Sk c log2 (S0k c ), (8) Ljoint(Y, Y0, S, S0) =↵ ⇥ Lcls(Y, Y0)+ (1 ↵) ⇥ Lseg(S, S0), (9) where ↵ is a hyper-parameter which defines the contribution of both loss functions in the final loss. Similar to patch clas- sifier, the loss function (Ljoint ) is minimized with RMSprop optimizer [38]. H. Training Strategies We trained the proposed framework in four different ways with varying ability to capture the spatial context. First, the proposed framework is trained without attention block and by minimizing the Lcls(·) loss only. This configuration is 7 xity. The performance of these classifiers for CRC g is reported in Table III. Although, the performance classifiers is comparable, MobileNet shows superior mance with highest mean accuracy. On the other hand, on classifier shows consistent performance across three with the lowest standard deviation (Std.). TABLE III ACCURACY COMPARISON OF FOUR PATCH CLASSIFIERS. Network Fold-1 Fold-2 Fold-3 Mean Std. ResNet50 93.48 93.62 89.13 92.08 2.08 MobileNet 93.48 95.74 89.13 92.78 2.74 nception-v3 95.65 91.49 86.96 91.37 3.55 Xception 93.48 91.49 91.30 92.09 0.98 TABLE V ROBUSTNESS ANALYSIS OF FEATURE EXTRACTORS ACROSS DIFFERENT METHODS. Methods ResNet50 (%) MobileNet(%) InceptionV3(%) Xception(%) RA-CNN 1 (Avg) 94.25±2.70 93.52±3.55 94.23±3.71 94.96±2.72 RA-CNN 1 (Max) 93.52±1.87 93.51±3.10 94.23±2.07 93.54±3.03 RA-CNN 2 (Avg) 92.08±2.08 93.52±1.78 94.96±2.72 94.96±2.72 RA-CNN 2 (Max) 95.68±3.55 93.52±3.55 92.80±2.72 93.54±3.03 RA-CNN 3 (Avg) 93.51±3.10 94.25±2.70 95.68±1.78 95.68±3.55 RA-CNN 3 (Max) 94.23±2.07 92.82±2.01 94.25±2.70 94.96±2.72 Rank-sum 12 12 10 8 TABLE VI COMPARISON FOR DIFFERENT TRAINING STRATEGIES WITH XCEPTION AS FEATURE EXTRACTOR. Feature Standard Weighted Auxiliary Attention Xception - Max 94.01 94.49 94.73 95.21 Xception - Avg 95.20 94.72 94.72 94.00 Mean 94.61 94.60 94.72 94.61 3 LR-CNN RA-CNN 4 er L3⇥3 p is also The CB, B3 , is •, ✓B2 3 ) F0, ✓B4 3 )] •, ✓B6 3 )] 7 3 )], (4) ent convolution e context block rning ned by the LR- learning from compared to ed set of three or B3 ) is used ned in section global average er to make the sses. The final X is computed ! Ls(•), (5) context blocks spectively. The ith categorical hich is defined Y 0k c ), (6) learn in a multi-task setting, where the coarse segmentation- map guides the network to improve the individual patch based feature classification in addition to the prediction of the input image. This leads to a network with improved classifica- tion performance since it is minimizing both segmentation and classification loss simultaneously. The segmentation-map based loss function (Lseg ) and joint loss function (Ljoint ) are defined as: Lseg(S, S0) = 1 K K X k=1 C X c=1 Sk c log2 (S0k c ), (8) Ljoint(Y, Y0, S, S0) =↵ ⇥ Lcls(Y, Y0)+ (1 ↵) ⇥ Lseg(S, S0), (9) where ↵ is a hyper-parameter which defines the contribution of both loss functions in the final loss. Similar to patch clas- sifier, the loss function (Ljoint ) is minimized with RMSprop optimizer [38]. H. Training Strategies We trained the proposed framework in four different ways with varying ability to capture the spatial context. First, the proposed framework is trained without attention block and by minimizing the Lcls(·) loss only. This configuration is represented by solid line blocks in Fig 2. Second, the same configuration as first but trained with a sample-based weighted loss function, Lwgt(·), which give more weight to the image patches with relatively less region of interest (glandular region) as compared to the background. The weight of an image and Lwgt(·) are defined as follow, Wk = ( 1 Rk roi , if Rk roi > ↵ 1 ↵ , otherwise (10) Lwgt(Y, Y0) = 1 K K X k=1 C X c=1 WkY k c log2 (Y 0k c ), (11) α=0.1 A. Network Input The input to our framework is an image (Xk) from a dataset, D = {Xk, Y k, Sk; k = 1, . . . K}, of large high resolution images which consists of K images with corresponding labels Y k 2 {1, . . . , C} for classification into C classes and coarse patch level segmentation masks Sk 2 {1, . . . , C} for multi- task learning. Each image is divided into M ⇥ N patches of same size where xk ij and yk ij represent the ijth patch of kth image and its corresponding label, respectively. We used a patch dataset, d = {(xk ij , yk ij ), | xk ij 2 Xk, yk ij 2 Y k}, which consists of patches and their corresponding labels for pre- training of LR-CNN. B. Local Representation Learning First part of the proposed framework encodes an input image Xk into a feature-cube Fk. All the input images are processed through the LR-CNN in a patch based manner. The proposed framework is flexible enough to use any state-of- the-art image classifier as a LR-CNN such as ResNet50 [33], MobileNet [34], Inception [35], or Xception [36]. This flex- ibility also enables it to use pre-trained weights in case of a limited dataset. Moreover, it is possible to train the LR-CNN independently before plugging it into the proposed framework, enabling it to learn meaningful representation [37] which leads to early convergence of the context-aware learning part of the framework. each value in the feature-cube. Hadamard product is taken between the weights and input feature-cube to increase the impact of more important areas of an image in label prediction and vice-versa. The weighted feature-cube F0 is defined as: F0 = L1⇥1 c (F, ✓c) ! Ls(•) ⌦ F, (2) where L1⇥1 c and ✓c represent the 1 ⇥ 1 convolution layer and its parameters, respectively. Ls denotes the softmax layer and the operator ⌦ is used to represent Hadamard product. E. Context Blocks Since the LR-CNN is used to encode the important patch- based image representation into a feature-cube, therefore the main aim of the context block (CB) is to learn the spatial context within the feature cube. The CB learns the relation between the features of the image patches considering their spatial location. We propose three different CB architectures, each with different complexity and capability to capture the context information. First CB, B1(·), is comprised of a 3 ⇥ 3 convolution layer followed by ReLU activation and batch normalization. Second CB, B2(·), uses residual block [33] architecture with two different filter sizes. It consists of three convolution layers each followed by batch normalization and ReLU activation. The first and last layers are with 1 ⇥ 1 convolution filter to squeeze and expand the feature depth. The ! L3⇥3 c (•, ✓B3 3 )] [L1⇥1 c (F0, ✓B4 3 )] [L1⇥1 c (F0, ✓B5 3 ) ! L3⇥3 c (•, ✓B6 3 )] [L3⇥3 p (F0) ! L1⇥1 c (•, ✓B7 3 )], (4) where ✓B1 3 to ✓B7 3 are the parameters of different convolution layers and ✓B3 represents parameter of the whole context block for the sake of notational simplicity. F. Representation Aggregation for Context Learning The local representation tissue has been learned by the LR- CNN. Therefore, the task of spatial context learning from feature-cube is relatively less challenging as compared to context learning from the raw image. A cascaded set of three context blocks (C(·)) of the same type (B1 ,B2 , or B3 ) is used in RA-CNN. These context blocks are explained in section III-E. The output of C(·) is followed by a global average pooling, a fully connected, and a softmax layer to make the final prediction in the required number of classes. The final prediction Y0 from the features of input images X is computed as: Y0 = C(F0, ✓C) ! Lg p (•) ! Lf (•, ✓f0 ) ! Ls(•), (5) where ✓C and ✓f0 represent the parameters of all context blocks and the fully connected layer in RA-CNN, respectively. The proposed framework is trained end-to-end with categorical cross-entropy loss based cost function Lcls(·) which is defined as: Lcls(Y, Y0) = 1 K K X k=1 C X c=1 Y k c log2 (Y 0k c ), (6) where Y k c and Y 0k c are the ground truth and predicted proba- bilities of kth image for cth class. and classification loss simultaneously. The segmentation based loss function (Lseg ) and joint loss function (Ljoint defined as: Lseg(S, S0) = 1 K K X k=1 C X c=1 Sk c log2 (S0k c ), Ljoint(Y, Y0, S, S0) =↵ ⇥ Lcls(Y, Y0)+ (1 ↵) ⇥ Lseg(S, S0), where ↵ is a hyper-parameter which defines the contrib of both loss functions in the final loss. Similar to patch sifier, the loss function (Ljoint ) is minimized with RMS optimizer [38]. H. Training Strategies We trained the proposed framework in four different with varying ability to capture the spatial context. First proposed framework is trained without attention block by minimizing the Lcls(·) loss only. This configuratio represented by solid line blocks in Fig 2. Second, the configuration as first but trained with a sample-based weig loss function, Lwgt(·), which give more weight to the im patches with relatively less region of interest (glandular reg as compared to the background. The weight of an image Lwgt(·) are defined as follow, Wk = ( 1 Rk roi , if Rk roi > ↵ 1 ↵ , otherwise Lwgt(Y, Y0) = 1 K K X k=1 C X c=1 WkY k c log2 (Y 0k c ), where Rk roi and Wk represent the ratio of the regio interest and weight of the kth image. The ↵ is the threshold, selected empirically as 0.10, sets the upper Xception RA-CNN3
  7. 最終結果とまとめ § 著者らが提唱したcontext-aware deep neural networkでは,病理診断に必要な形態学的特徴を様々な倍率の画像か ら効率的に学習することで標準的なパッチサイズの64倍の量をたった10%の計算時間の増加で扱うことができ, ⼤腸癌のグレード分類において99.28%(2クラス),95.7%(3クラス)の⾼い精度を⽰した. § ⼤きなコンテクスト情報が必要なWSI診断に適しており,他のがん腫への応⽤に期待したい

    9 . 4. Visual results of CRC grading are shown for patch classifier, existing context, and the proposed method on an image of size 1, 792 ⇥ 1, 792. The de size for context networks is equal to the size of patch (224 ⇥ 224) used for patch classifier. Green, blue and red colors of overlaid rectangular boxes w the normal, low and high-grade predictions respectively, whereas empty box areas represent non-glandular/background regions. TABLE VIII RANK-SUM BASED COMPARISON AS MEASURED BY THE F1-MEASURE ON BREAST DATASET. Methods Classes A B C D E F G H I J Proposed Normal 0.501 0.468 0.523 0.513 0.509 0.603 0.573 0.252 0.241 0.323 0.643 Benign 0.453 0.468 0.482 0.444 0.410 0.369 0.423 0.489 0.333 0.437 0.511 InSitu 0.468 0.476 0.486 0.533 0.615 0.614 0.581 0.286 0.311 0.452 0.362 Invasive 0.401 0.477 0.430 0.54 0.557 0.548 0.576 0.520 0.446 0.580 0.576 Rank-sum 23 22 22 20 16 16 17 20 24 18 10 VI. CONCLUSION model is presented and compared with existing approaches in Fig. 3. (Left) Results of 24 experiments using best performing local representation features (Xception). (Right) Legend represents the feature pooling type, context-aware net and training strategies used for the experiments. Baseline accuracy is the accuracy of standard patch based Xception classifier. robustness required for multi-class grading of CRC images (see Table VII). Our best performing context-aware network, RA-CNN 3, with Xception based LR-CNN and attention based training method achieved superior performance as compared to the BAM based methods. We also compared our method with a set of context- aware approaches explored in a systemic study on context- aware learning by Sirinukunwattana et al. [29]. Ten different approaches (A-J) were considered to capture contextual infor- mation. For a detailed description of these methods please see Figure 2 of [29]. It can be observed that these approaches have significant overlap with each other. some of these approaches differ only in term of input patch resolution, as some of these used shared weights for different networks. First four (A- D) approaches try to capture the context by down-sampling the high-resolution images at different magnification levels e.g. 20⇥, 10⇥, 5⇥, and 2.5⇥. Channel-wise concatenated four multi-resolution images as an input of a CNN are consider in approach E. However, approaches F and I concatenate the CNN features of four multi-resolution input images in vector form before prediction. The full image at 20⇥ magnification is used as an input in approach H. This approach is not feasible in case of an image with very large spatial dimensions due to memory constraints. Approaches G and J use LSTM to capture the context from the CNN features of four multi-resolution input images. The code of method G is publicly available by the authors of [29] and we use that code to retrain the LSTM based method on CRC dataset for a fair comparison. TABLE VII COMPARISON WITH STATE-OF-THE-ART ON COLORECTAL DATASET. Method Binary (%) Three-class (%) BAM - 1 [11] 95.70 - 2.10 87.79 - 2.32 BAM - 2 [11] 97.12 - 1.27 90.66 - 2.45 Context - G [29] 96.44 - 3.61 89.96 - 3.54 ResNet50 [33] 98.57 - 1.01 92.08 - 2.08 MobileNet [34] 97.83 - 1.77 92.78 - 2.74 InceptionV3 [35] 98.57 - 1.01 91.37 - 3.55 Xception [36] 98.58 - 2.01 92.09 - 0.98 Proposed 99.28 - 1.25 95.70 - 3.04 irregular for any given image due to the lake of contex- tual information. The predictions of Context-G are relatively smooth but it predicts the wrong label for the low-grade image which might be due to the use of a low-resolution image for context learning. However, the proposed method predictions are smooth and consistent with the ground truth labels. We retrained our method on breast cancer dataset to make a direct comparison to all the context-aware approaches pre- sented in [29]. Sirinukunwattana et al. used a customized network as feature extractor which contains 5 convolution layers with 4⇥4 filter and each layer followed by batch-norm and leaky-ReLU activation. To highlight the significance of RA-CNN, we adopted their network for LR-CNN and trained our RA-CNN 3 with standard training strategy in end-to-end training manner. Moreover, the same experimental setup (as in ⼤腸癌セットにおけるコンテクストレベルのaccuracy⽐較 乳癌セットにおけるF1 scoreの⽐較 Fig. 4. Visual results of CRC grading are shown for patch classifier, existing context, and the proposed method on an image of size 1, 7 stride size for context networks is equal to the size of patch (224 ⇥ 224) used for patch classifier. Green, blue and red colors of overlaid show the normal, low and high-grade predictions respectively, whereas empty box areas represent non-glandular/background regions. TABLE VIII RANK-SUM BASED COMPARISON AS MEASURED BY THE F1-MEASURE ON BREAST DATASET. Methods Classes A B C D E F G H I J Proposed Normal 0.501 0.468 0.523 0.513 0.509 0.603 0.573 0.252 0.241 0.323 0.643 Benign 0.453 0.468 0.482 0.444 0.410 0.369 0.423 0.489 0.333 0.437 0.511 InSitu 0.468 0.476 0.486 0.533 0.615 0.614 0.581 0.286 0.311 0.452 0.362 Invasive 0.401 0.477 0.430 0.54 0.557 0.548 0.576 0.520 0.446 0.580 0.576 Rank-sum 23 22 22 20 16 16 17 20 24 18 10 VI. CONCLUSION In this paper, we present a novel context-aware deep neural network for cancer grading, which is able to incorporate 64 times larger context than standard CNN based patch classi- fiers. The proposed network is well-suited for CRC grading task which relies on recognizing abnormalities in glandular structures. These clinically significant structures vary in size and shape that cannot be captured efficiently with standard patch classifiers due to computational and memory constraints. model is presented and compared with existing the same evaluation setting. The qualitative an results demonstrate that our method outperfor based classification methodologies, the domain niques and existing context-based methods. Th is suitable for cancer analysis which requires la information in the histology images. This inc grading in prostate cancer and tumor growth fication in lung cancer. Moreover, this approa be extended to perform downstream analysis a ⻘; normal 緑; low grade ⾚; high grade blank; background/no gland
  8. Supplementary Document by the author • https://www.dropbox.com/s/f6t4unyqf6waf3m/Supplementary%20D ocument.pdf?dl=0 • Video

    demos High Grade Image http://bit.ly/2UsGywg Low Grade Image http://bit.ly/2IBYiTv Normal Image http://bit.ly/2DijOZP • Online demo https://colab.research.google.com/drive/1j5HP0WbJVj5z9IQpI5FRfW N4Ww2H4RJC