Infuse AI to Your Enterprise Yonghua LIN, IBM Research IBM - - PowerPoint PPT Presentation

Infuse AI to Your Enterprise

Yonghua LIN, IBM Research

IBM Distinguished Engineer Leader of AI System Research

IBM Opens New Era of Artificial Intelligence

2

IBM Watson Watson won humans in Jeopardy

Quantum Computing Neural Chips In-memory Computing Deep Learning Systems

Solution Supporting Technologies

Transportation Cognitive Medical Finance & Insurance Smarter City Media & Entertainment Cognitive Retailer Automobile Manufacture AI Cloud Computing AI Vision, Acoustic, Language, Conversation Deep Learning & Machine Learning

Global research teams formed the base to drive AI innovation in IBM

3

• IBM’s 2016 patent output features more than 2,700 patents for

inventions related to AI, cognitive computing and cloud computing

• IBM Research has deep understanding on vision technology and

published hundreds of high quality papers on top AI and computer vision conferences, such as CVPR, AAAI, NIPS, etc..

DEMO

This is the first AI Highlight for TV entertainment show. Key words: Multi-model learning, AI for Media

4

AI-made Film / Cognitive Highlight

• To generate the highlight film for video will consume lots of time and human effort.
• With AI technology, machine could understand each video frame and generate a

highlight film automatically.

Scene detection Emotion identification Face detection Audio identification Speech-to Text Action detection

…

Multi-task models Generate index for video One hour video: 80K entries and 100MB table

Inference on video streams Create highlight films by machine

Highlight strategies New films

Challenges for enterprises going into AI DATA

6

TALENTS

The hour glass for deep learning

To build a team with deep learning expertise : 2 months ~ 1 year To prepare massive training data : ~ 10 man months To train a new model : 1 hour ~ weeks To give an AI inference result : < 1s

Steps for AI Deep Learning Development

Usually, developers need following steps to develop a DNN model and make it usable for application

Define training task Prepare training Data Data Pre- processing DNN Model selection Configure the training hyper- parameter DNN Model Training

Start

Package the new DNN model together with preprocessing into inference proc.

Application development with inference API

DL training framework preparation

• No experience on DNN design and develop
• No experience on computer vision
• No experience on how to build a platform to support enterprise scale deep learning,

including data preparation, training, and inference

Most of enterprises are facing the challenges …

Define training task Prepare training Data Data Pre- processing DNN Model selection Configure the training hyper- parameter DNN Model Training

Start

Package the new DNN model together with preprocessing into inference proc. DL training framework preparation Data Pre- processing DNN Model selection Configure the training hyper- parameter DNN Model Training DL training framework preparation Package the new DNN model together with preprocessing into inference proc.

Steps automatically done by PowerAI Vision

We can help deep learning for Vision easier – PowerAI Vision

Deep knowledges of ML/DL and computer vision have been embedded into PowerAI Vision. Easily develop AI application for Computer Vision !!

User could use the deployed API for visual recognition

Example 1: AI for Product Quality Inspection (Manufacture)

Inspect images of photoresist openings after having been exposed and developed (光刻是通过一系列生产步骤,将晶圆表面薄膜的特定部分去除

的工艺。被广泛用于集成电路的生产流程。显影检查需要人工检验不合格的晶 圆，以便返工重新曝光、显影。) 显影检查：图形尺寸的偏差、光刻胶的污染、空洞、划伤，以及污点等。

With PowerAI Vision, the manufacture could quickly build the auto defect inspection capability :

• Data import：１５min.
• Data labeling：５min.
• AI Vision training：１０min.

Accuracy: 94.5%

PowerAI Vision : End-to-end DL Development and Operation for Computer Vision

Data Lake & Data Stores Distributed Computing ML & DL Libraries & Frameworks

PowerAI Accelerated Servers Storage

Inference API deployment for Cloud Video and Image Labeling Data Preprocessing Self-defined Training with visualized monitoring Custom Learning

Training

PowerAI Vision: Development Pipeline

Data

Inference accelerator generation for edge

Inference

Testing & Measurement

PowerAI Software Stacks

Custom Learning for Image and Video

Custom learning for classification Custom learning for object detection Image and Video Analysis

Focus of PowerAI Vision

• AI for AI

Transfer Learning for Learning from Small Data Set

14

• In lots of industry scenarios, we don’t have huge data set.
• PowerAI Vision applied the optimized Transfer Learning technology for custom

learning from small data set.

Good base model Small data set from User Base models supported by PowerAI Vision

• 1. Select

data scenario (base model)

• 2. Add user’s

training data

• 3. One-click

to start training

Small data set, better accuracy, faster training

Data Augmentation for Learning from Small Data Set

• Data Augmentation can enhance the classification accuracy and reduce overfitting for small

datasets

• Data Augmentation functions has been available on PowerAI Vision

15

Fig.1 User could “one-click” and select different data augmentation methods Fig.2 Data augmentation could improve the accuracy significantly

Medical image analysis for cerebral hemorrhage (脑出血) (Original data: 157 pic.)

Accuracy: 97.9%

DL for DL: Learning to optimize parameters for visual analysis

16

• Through machine learning, PowerAI Vision will automatically tune parameters to achieve good

accuracy for different training cases defined by users.

• In the following test case, our auto-tuning DL network could outperform the fix manual

configuration (default) by >6%. And it could achieve the same accuracy (e.g. 90%) with much less training time (e.g. <1/3).

• Fig. 1 Performance comparison for object detection

Auto-tune

18 parameters have been tuned, including

• Caffe training parameters
• Neural network parameters
• Object detection

parameters.

Test data set: object detection for helmet and safety vest

Semi auto-labeling : Reduce the time for data annotation

• Semi – auto labeling : To use AI technology for releasing most of human work

for labeling (10x ~ 50x)

17

Manually label small data set System will learn the objects for labeling Auto – labeling by machines Human review and adjust

DEMO

This is a demo for retail scenario. In this demo, audience can see how to use PowerAI Vision to learn and detect shopping cart easily. Key words: custom learning for object detection, auto-labeling

18

PowerAI Vision for Retail Business

• 1. Real-time understand your customer and make

recommendation

• Age and gender
• What she wears
• Styles: office lady,

fashion, etc.

• Shopping with kids,

friend, or parents

• …
• 3. Shelves management
• Goods

counting on the shelves

• Goods

distribution on the shelves

• 2. Client flow analysis
• Hot area analysis
• Client behavior

analysis

• Queue length

monitoring

• 4. Risk detection for security
• Fighting

detection

• Wearing mask

detection

• Occupation in

sensitive area (e.g. exit path)

• 5. Anti-Lost
• Goods

tracking at the counter

• illegal

behavior analysis

企业级的人工智能系统需要极高的综合技术能力：车载辅助驾驶 系统

汽车企业需要构建端到端的深度学习开发平台，支持从数据中心到车载系统的 AI 研发

20

PowerAI Vision 深度学习开发

训练数据集 预处理 一键启动深 度学习训练 监测训练过程 标注视频训 练数据 自动化生成FPGA 加速器软件 FPGA嵌入式系统FPGA (Znyq70xx series)

小目标

实时低时延：每帧9ms 超小目标检测 : 12x12像素 多种天气及光照条件 单目高精准测距

路测效果

PowerAI Vision : Infuse AI into Your Enterprise DATA

21

TALENTS Innovation with AI for AI