Upgrade to Pro — share decks privately, control downloads, hide ads and more …

卷积神经网络

Yuchu Luo
November 21, 2017
320

 卷积神经网络

Yuchu Luo

November 21, 2017
Tweet

Transcript

  1. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. 2017. ImageNet

    classification with deep convolutional neural networks. Commun. ACM 60, 6 (May 2017) 分类(Classification)
  2. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. 2017. ImageNet

    classification with deep convolutional neural networks. Commun. ACM 60, 6 (May 2017) 检索(Retrieval)
  3. person 1.0 person 1.0 person 0.9 person 0.9 person 0.4

    person 1.0 person 1.0 person 1.0 bottle .96 bottle .99 bottle .99 table .96 Kaiming He, Georgia Gkioxari, Piotr Dolla ́r and Ross Girshick (2017). Mask R-CNN. CoRR, abs/1703.06870, . 检测 分类 分割 (Detection, classification, segmentation)
  4. Kaiming He, Georgia Gkioxari, Piotr Dolla ́r and Ross Girshick

    (2017). Mask R-CNN. CoRR, abs/1703.06870, . 检测 分类 分割 (Detection, classification, segmentation)
  5. Kaiming He, Georgia Gkioxari, Piotr Dolla ́r and Ross Girshick

    (2017). Mask R-CNN. CoRR, abs/1703.06870, .
  6. Dongdong Chen and (2017). StyleBank: An Explicit Representation for Neural

    Image Style Transfer. CoRR, abs/1703.09210, . ⻛风格迁移(Style transfer)
  7. 1 1 3 -1 0 2 1 -1 0 ⼀一维卷积示例例

    -2 -1 1 0 滤波器器(3x1) 1 -2 1 3 -1 -1 0 = x + x + x 2 1 1 1 -1 0 = x + x + x 0 填充(Padding) 步⻓长(Stride)
  8. Olah, et al., "Feature Visualization", Distill, 2017. Input output 3a

    3b 4a 4b 4c 4d 4e 5a 5b GoogLeNet 可视化 演示以及论⽂文地址:https://distill.pub/2017/feature-visualization/ Feature visualization
  9. GoogLeNet 可视化 Feature visualization Olah, et al., "Feature Visualization", Distill,

    2017. 从随机 noise 开始, 优化 (optimize) 图⽚片以激活特定的神经元 Step 0 Step 4 Step 48 Step 2048
  10. 可 视 化 效 果 激 活 正 样 本

    Olah, et al., "Feature Visualization", Distill, 2017.
  11. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 检测出了了简单的纹理理信息(texture),且都是局部(local)纹理理 Olah, et al., "Feature Visualization", Distill, 2017.
  12. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 纹理理开始变得复杂,仍然都是局部纹理理 Olah, et al., "Feature Visualization", Distill, 2017.
  13. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 复杂度显著提升,出现了了复杂的模式(pattern) Olah, et al., "Feature Visualization", Distill, 2017. 书架 狗眼睛 ⽂文本,柳柳钉 ⻦鸟
  14. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 可以辨别出部分物体了了,更更多的 context 开始出现 Olah, et al., "Feature Visualization", Distill, 2017. 建筑结构 蓬松的绳 树 台球
  15. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b ⽹网络开始响应具体的对象(这也许是最值得探索的⼀一层) Olah, et al., "Feature Visualization", Distill, 2017. 棕榈树 轮⼦子 带项圈的狗 房⼦子
  16. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 出现更更复杂的概念,开始看到神经元对多个不不相关的概念产⽣生反应 Olah, et al., "Feature Visualization", Distill, 2017. 狗嘴 灵⻓长⽬目动物 蛇头 餐厅的菜肴
  17. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 区分特定的物种,对多个视觉相关的概念产⽣生响应 Olah, et al., "Feature Visualization", Distill, 2017. ⻳龟壳 冰淇淋淋 & ⾯面包 猫的⽪皮肤 宽边帽
  18. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 可视化开始变得抽象,但所包含的语义信息还是较为具体 Olah, et al., "Feature Visualization", Distill, 2017. 糖果 球 铜管乐器器 交通灯
  19. Input output 3a 3b 4a 4b 4c 4d 4e 5a

    5b 像是没有意义的⼤大杂烩,神经元似乎不不再对应特别的语义概念 Olah, et al., "Feature Visualization", Distill, 2017.
  20. l l l l l ⻛风格图⽚片 内容图⽚片 ⻛风格 Loss +

    内容 Loss 底层纹理理特征 深层语义特征 优化⽬目标: Gatys, Ecker, and Bethge, “Image style transfer using convolutional neural networks”, CVPR 2016 Figure copyright Justin Johnson, 2015.
  21. l l l l l Gatys, Ecker, and Bethge, “Image

    style transfer using convolutional neural networks”, CVPR 2016 Figure copyright Justin Johnson, 2015.
  22. High-level Task: 视觉推理理 Q: Are there an equal number of

    large things and metal spheres? Q: What size is the cylinder that is left of the brown metal thing that is left of the big sphere? Q: How many objects are either small cylinders or red things? AI 不不仅需要识别出图像中的物体, 还需要有推理理、判断能⼒力力 Justin Johnson and (2016). CLEVR: A Diagnostic Dataset for Compositional Language and Elementary. CoRR, abs/1612.06890, .
  23. High-level Task: 视觉会话 Abhishek Das and (2017). Learning Cooperative Visual

    Dialog Agents with Deep Reinforcement. CoRR, abs/1703.06585, .
  24. 我 卷积神经⽹网络(CNN)与⾃自然语⾔言处理理(NLP) 有 ⼀一头 ⼩小 ⽑毛驴 5 0 2 1

    3 1 4 1 0 1 0 2 4 3 3 0 2 1 1 8 2 7 3 1 5 0 1 2 3 5 Yoon Kim (2014). Convolutional Neural Networks for Sentence Classification. CoRR, abs/1408.5882, . 语⾔言具有局部相关性
  25. Gehring, Jonas,, Auli, Michael, Grangier, David, Yarats, Denis & Dauphin,

    Yann N (2017). "Convolutional Sequence to Sequence Learning". ArXiv e-prints https://github.com/facebookresearch/fairseq 使⽤用卷积神经⽹网络(CNN)的机器器翻译模型
  26. 卷积神经⽹网络(CNN) 与语⾳音识别(Speech Recognition) Eric Battenberg and (2017). Exploring Neural Transducers

    for End-to-End Speech Recognition. CoRR, abs/1707.07413, . Tara N. Sainath, Ron J. Weiss, Andrew Senior, Kevin W. Wilson, Oriol Vinyals, “Learning the Speech Front-endWith Raw Waveform CLDNNs,” In INTERPSEECH 2015 输⼊入声⾳音原始波形 输出识别结果