41
Pushing the boundaries of ML A case of Simplification & Scale Magnus Hyttsten The AI Summit San Francisco

The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

Page 1: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Pushing the boundaries of MLA case of Simplification & Scale

Magnus Hyttsten

The AI SummitSan Francisco

Page 2: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes
Page 3: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Guinea Pig

Meet Robin

Page 4: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

“The network performed similarly to senior orthopedic surgeons when presented with images at the same resolution as the network.”

www.tandfonline.com/doi/full/10.1080/17453674.2017.1344459

RadiologyOphthalmology

0.95Algorithm Ophthalmologist (median)

0.91

Page 5: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Pathology

Page 6: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

neural (GNMT)

phrase-based (PBMT)

English >

Spanish

English >

French

English >

Chinese

Spanish >

English

French >

English

Chinese >

English

Translation model

Tran

slat

ion

qual

ity

0

1

2

3

4

5

6

human

perfect translation

Closes gap between old system and human-quality translation from 58% to 87%

Enables better communication across the world

Neural Machine Translation

Page 7: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Robotics - Discovery Algorithms

Page 8: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

ImageNet

Alaskan Malamute Siberian Husky

Page 9: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes
Page 10: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes
Page 11: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

https://cloud.google.com/blog/big-data/2016/08/how-a-japanese-cucumber-farmer-is-using-deep-learning-and-tensorflow

Page 12: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Cucumber sorter

From: http://workpiles.com/2016/02/tensorflow-cnn-cucumber/

Page 13: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes
Page 14: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Data, Data, Data

Compute, Compute, Compute

Page 15: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Data, Data, Data

Compute, Compute, Compute

Humans, Humans, Humans

Page 16: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Improving Inception and Image Classification in TensorFlowresearch.googleblog.com/2016/08/improving-inception-and-image.html

How long did it take for a Human to construct this?

Page 17: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

AM!!!

Page 18: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Current:Solution = ML expertise + data + computation

Page 19: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Current:Solution = ML expertise + data + computation

Can we turn this into:Solution = data + 100X computation

Page 20: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Current:Solution = ML expertise + data + computation

Can we turn this into:Solution = data + 100X computation

???Can We Learn How To Teach Machines To Learn

Page 21: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

CIFAR-10

Page 22: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Learning Transferable Architectures for Scalable Image Recognition, Barret Zoph, Vijay Vasudevan, Jonathon Shlens and Quoc Le, https://arxiv.org/abs/1707.07012

ImageNet

Page 23: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Large-Scale Evolution of Image ClassifiersEsteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Jie Tan, Quoc Le, Alex Kurakinhttps://arxiv.org/abs/1703.01041

Page 24: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Using RL to Distribute Training to

GPUs

Device Placement Optimization with Reinforcement Learning, Azalia Mirhoseini, Hieu Pham, Quoc V. Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi,Samy Bengio, Jeff Deanhttps://arxiv.org/abs/1706.04972

Page 25: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Using RL to Distribute Training to

GPUs

Page 26: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Growing use of deep learning @ Google

Page 27: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Making this Happen

A Case of Simplification

Page 28: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

● An open-source machine learning framework for everyone

● Fast, flexible, and production-ready on all major platforms

● Able to scale from research to production

TensorFlow is Strategic to Google

Page 29: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

TensorFlow Distributed Execution Engine

CPU GPU Android iOS ...

C++ FrontendPython Frontend ...

Estimator

tf.layers

tf.keras

Canned Estimators

Data

sets

Page 30: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

https://goo.gl/Ujm2Ep

Getting Started

Page 31: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes
Page 32: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Demo

Cloud Video Intelligence(animated .gif here)

Page 33: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Making this Happen

A Case of Scale

Page 34: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Scaling to multiple platforms

iOS

CPU GPU

Android

Cloud TPU

Embedded Devices

Page 35: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Scaling Training Speed and

CPU GPU TPU

Page 36: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Tensor Processing Unit (TPU) v2

Page 37: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Tensor Processing Unit (aka TPU) v2Designed for neural net training and inference

● 180 teraflops of computation, 64 GB of HBM memory, 2400 GB/s memory bandwidth

● Designed to be connected together

Page 38: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

TPU Pod - 11.5 petaflops

Page 39: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Pushing the boundaries of ML

● Machine Learning & Artificial IntelligenceProven & Ready for Business

A case of simplification & scale

Page 40: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Pushing the boundaries of ML

● Machine Learning & Artificial IntelligenceProven & Ready for Business

● But Designing & Training ModelsTakes Time & Is Complex

A case of simplification & scale

Page 41: The AI Summit San Francisco Pushing the boundaries of ML A ...€¦ · Machine Learning & Artificial Intelligence Proven & Ready for Business But Designing & Training Models Takes

Pushing the boundaries of ML

● Machine Learning & Artificial IntelligenceProven & Ready for Business

● But Designing & Training ModelsTakes Time & Is Complex

● TensorFlow, TPUs, and Model AutomationUsed by Google to push the boundaries of ML

A case of simplification & scale