O'reilly AI Conference - Speaker Deck

Slide 1

Slide 1 text

San Francisco, CA 17 - 20 September 2017 Bramantyo Adrian

Slide 2

Slide 2 text

What “It is an AI Event that focus on real-world implementations.” GDP Labs Confidential

Slide 3

Slide 3 text

PUT AI TO WORK

Slide 4

Slide 4 text

1. Building Unbiased AI 2. Accelerating AI 3. Active Learning and Transfer Learning 4. Object Detection GDP Labs Confidential

Slide 5

Slide 5 text

Introduction + Internet Company IS NOT GDP Labs Confidential

Slide 6

Slide 6 text

Introduction + AI Company IS NOT GDP Labs Confidential

Slide 7

Slide 7 text

1. Building Unbiased AI 2. Accelerating AI 3. Active Learning and Transfer Learning 4. Object Detection GDP Labs Confidential

Slide 8

Slide 8 text

Diversity Crisis in AI “We’ve already seeing society’s racial and gender biases being encoded into software that uses AI when built by such a homogeneous group.” Building Unbiased AI GDP Labs Confidential

Slide 9

Slide 9 text

Building Unbiased AI GDP Labs Confidential

Slide 10

Slide 10 text

Building Unbiased AI GDP Labs Confidential

Slide 11

Slide 11 text

Goals Building Unbiased AI GDP Labs Confidential

Slide 12

Slide 12 text

Building Unbiased AI GDP Labs Confidential

Slide 13

Slide 13 text

Ignoring - Failing to max a product to different groups - Failing to attract potential users Building Unbiased AI GDP Labs Confidential

Slide 14

Slide 14 text

1. Building Unbiased AI 2. Accelerating AI 3. Active Learning and Transfer Learning 4. Object Detection GDP Labs Confidential

Slide 15

Slide 15 text

Accelerating AI GDP Labs Confidential

Slide 16

Slide 16 text

Manufacturing Automotive Healthcare Financial Transforming Many Industries Accelerating AI GDP Labs Confidential

Slide 17

Slide 17 text

CPU GPU FPGA ASIC Computational Substrate Accelerating AI GDP Labs Confidential

Slide 18

Slide 18 text

why Deep Learning ? Accelerating AI GDP Labs Confidential

Slide 19

Slide 19 text

Latest Model Not Always Relevant Model Accelerating AI GDP Labs Confidential

Slide 20

Slide 20 text

SSD model designed and optimized for PASCALVOC/MSCOCO dataset Know Your Model Provenance Accelerating AI GDP Labs Confidential

Slide 21

Slide 21 text

Smart hospital will generate over 3,000 GB PER DAY Self driving cars will generate over 4,000 GB PER DAY … EACH The average internet user will generate ~1.5 GB OF TRAFFIC PER DAY A connected plane will generate over 40,000 GB PER DAY A connected factory will generate over 1,000,000 GB PER DAY Invest in Data Manual annotations Training data Accelerating AI GDP Labs Confidential

Slide 22

Slide 22 text

1. Building Unbiased AI 2. Accelerating AI 3. Active Learning and Transfer Learning 4. Object Detection GDP Labs Confidential

Slide 23

Slide 23 text

Active and Transfer Learning Active Attain good learning performance without demanding too many labeled data GDP Labs Confidential

Slide 24

Slide 24 text

Motivation Active and Transfer Learning GDP Labs Confidential

Slide 25

Slide 25 text

Human Active AI Classifier Output Human Annotation Confident N ot Confident Active Learning Active and Transfer Learning GDP Labs Confidential

Slide 26

Slide 26 text

Transfer It’s very expensive to train deep learning Active and Transfer Learning GDP Labs Confidential

Slide 27

Slide 27 text

Transfer as feature extractor Active and Transfer Learning GDP Labs Confidential

Slide 28

Slide 28 text

Transfer as fine tuning Active and Transfer Learning GDP Labs Confidential

Slide 29

Slide 29 text

1. Building Unbiased AI 2. Accelerating AI 3. Active Learning and Transfer Learning 4. Object Detection GDP Labs Confidential

Slide 30

Slide 30 text

Computer Vision Object Detection GDP Labs Confidential

Slide 31

Slide 31 text

Traditional Machine Learning to Deep Learning Object Detection Feature Engineering challenges: - Human Scaling - Computational Scaling GDP Labs Confidential

Slide 32

Slide 32 text

Object Detection Feature Learning challenges: - Need large datasets - Model become complex Traditional Machine Learning to Deep Learning GDP Labs Confidential

Slide 33

Slide 33 text

Object Detection CIFAR, ImageNet PascalVOC MSCOCO KITTI GDP Labs Confidential

Slide 34

Slide 34 text

Convolutional Neural Network Object Detection GDP Labs Confidential

Slide 35

Slide 35 text

Object Detection Object Detection GDP Labs Confidential

Slide 36

Slide 36 text

Object Detection History Object Detection Region Proposed Approach Direct Classification Refined Classification GDP Labs Confidential

Slide 37

Slide 37 text

Region Proposed Approach Object Detection - Run through image to detect regions - Process regions extensively: - Bounding boxes - Predict class GDP Labs Confidential

Slide 38

Slide 38 text

Direct Classification Object Detection - BB and class are predicted directly with a single network GDP Labs Confidential

Slide 39

Slide 39 text

Refined Classification Object Detection - Add another branch that predict pixel-wise classes GDP Labs Confidential

Slide 40

Slide 40 text

Training Vision Model Challenges Object Detection GDP Labs Confidential

Slide 41

Slide 41 text

Training Vision Model Challenges Object Detection Batch size epochs Top-1 Accuracy Hardware Cost ($) Time 256 90 73.0% 1 DGX station 129,000 21h 8192 90 72.7% 1 DGX station 129,000 21h 8192 90 72.7% 32 DGX station 4.1 million 1h 32K 90 72.4% 512 KNLs 1.2 million 1h GDP Labs Confidential