Google’s Journey to AI-First “Hey Google, What is AI First?”
Chanchal Chatterjee November 13-14, 2019 Slides © Google with many contributors. Not for redistribution.
Google Cloud Platform Copyright Google 2013-2017 1 Era of AI
Observations
General AI? Goal System Environment
Actions
Google Cloud Platform Copyright Google 2013-2017 2 Deep learning
Geoff Hinton’s work fascinated our team at Google X
We thought he was onto something with his ingenious simplifications and engineering of artificial neural networks
Google Cloud Platform Copyright Google 2013-2017 3 Is this a cat or a dog?
Google Cloud Platform Copyright Google 2013-2017 4 2015
Google Cloud Platform Copyright Google 2013-2017 5 2017
Google Cloud Platform Copyright Google 2013-2017 6 A B
From Google Blog. 2018
Google Cloud Platform Copyright Google 2013-2017 7 Areas of Dramatic Improvement
Vision Language Translatio Speech Events n
Google Cloud Platform Copyright Google 2013-2017 8 Agenda
● Artificial Intelligence at Google ● Advancing Techniques ● Industry Applications
Google Cloud Platform Confidential & Proprietary 9 YouTube | WatchNext
Over 70 percent of video watch time now driven by algorithmic recommendations
Aggregate watch time on home page 20 times larger over 3 years - The Verge
Google Cloud Platform Copyright Google 2013-2017 10 RankBrain (a deep neural network for search ranking) improved performance significantly #3 Search signal for Search ranking, out machine learning for search engines of hundreds* #1 improvement to ranking quality in 2+ years* *Google Internal Data
Google Cloud Platform Copyright Google 2013-2017 11 Self-driving cars
Confidential & Proprietary 40 percent reduction in cooling energy
Servers
Servers
Facilities Facilities
Typical datacenter Google datacenter electricity usage electricity usage
Google Cloud Platform Copyright Google 2013-2017 13 Google Differentiates with AI
Assistant Search YouTube Query understanding Query understanding Video recommendations Conversation Search ranking Better thumbnails
Cloud Home Translate Cloud ML APIs Speech recognition text and speech TPU Conversation translations
Clips Gmail Photos smart image capture Smart Reply Photos / video search Spam classification Auto- smile montage
Android Cardboard Drive Keyboard input Image stitching Quick Access (also in iOS)
Maps Ads Play Street View images Richer Text Ads App clustering Parsing Local Search Video summarization Music recommendations Advancing Techniques
Google Cloud Platform Confidential & Proprietary 15 Data is the Foundation for AI
TensorFlow
Open Hadoop HBASE Crunch Drill Source Beam
Map Flume Google GFS BigTable Dremel Spanner Millwheel Dataflow TensorFlow Papers Reduce Java
Google Cloud Products BigQuery Pub/Sub Dataflow Bigtable ML
2002 2004 2005 2006 2008 2010 2012 2014 2015 2016
Confidential & Proprietary Google Cloud Platform Copyright Google 2013-2017 16 The most common ML models at Google are models that operate on structured data
Type of network # of network layers # of weights % of deployed models
MLP0 5 20M 61% MLP1 4 5M
LSTM0 58 52M 29% LSTM1 56 34M
CNN0 16 8M 5% CNN1 89 100M https://cloud.google.com/blog/big-data/2017/05/an-in-depth-look-at-googles-first-tensor-processing-unit-tpu
Google Cloud Platform Confidential & Proprietary 17 TFX: TensorFlow-Based ML platform at Google
Integrated Frontend for Job Management, Monitoring, Debugging, Data/Model/Evaluation Visualization
Shared Configuration Framework and Job Orchestration
Tuner
Data Data Data Data Model Evaluation Trainer Serving Logging Ingestion Analysis Transformation Analysis and Validation
Shared Utilities for Garbage Collection, Data Access Controls
Pipeline Storage
Confidential & Proprietary 18 Embeddings: learn signals
0.1 0.5 Learned -0.4 Embedding 128-D ... Function
0.02 ● From input images to low dimensional feature vectors ● Images of the “same class” are close together in the embedding space. ● Works for many data types (e.g. images, words, electrophysiology, transcription profiles) Text comprehension Berlin - Germany + Russia = Moscow take - took + drew = draw
Spain Italy Madrid Rome Germany fallen Berlin fall Turkey drawn given Ankara draw give Russia fell Moscow taken Canada Ottawa drew take Japan Tokyo gave Vietnam Hanoi took China Beijing
Solve analogies with vector arithmetic!
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean.Distributed Representations of Words and Phrases and their Compositionality. NIPS, 2013. YouTube Recommendations - 2016
Source: Deep Neural Networks for YouTube Recommendations
Google Cloud Platform Confidential & Proprietary 21 AI does AI Systematic exploration of the model space, using the techniques finessed in AlphaGo, yields super-human performance in AI network design
Source: Google Blog.
Google Cloud Platform Confidential & Proprietary 23 AutoML: How it works
A technology that can automatically create a Machine Learning Model
Source: Why AutoML Is Set To Become The Future Of Artificial Intelligence, Janakiram MSV. Google Cloud Platform Confidential & Proprietary 25 Industry Applications
Google Cloud Platform Confidential & Proprietary 26 To win in the future you need speed, scale and obsessive user focus
Cloud AI Mobile
Confidential & Proprietary Google Cloud Platform Copyright Google 2013-2017 27 AI-First Online Grocery
● 80x faster analytics for business agility ● 4x faster response to customer service claims with cloud Natural Language Processing ● 15x precision in fraud prediction with cloud Deep Learning ● Cloud ML models for customer lifecycle ● Processing and distributed ML for warehouse robots Business
Pixel Based Automation CPA, SA Sophisticated Ad Targeting Comprehensive Analytics
Opportunity to Improve ● Blended modeling ● Automation ● Integration
Data Analysis (Slow, Labor Intensive) Three ways to deliver AI capabilities
Infrastructure AI Building Blocks AI Solutions
Machine Learning AutoML Dialogflow Cloud Google Sheets Engine Vision Enterprise Edition Speech Differentiation Time to Value
Google Cloud Platform Copyright Google 2013-2017 3030 The Journey to AI-First
Realtime Big Data AI-First Digital
Business Impact
Maturity
Google Cloud Platform Copyright Google 2013-2017 31 Thank You
Google Cloud Platform Copyright Google 2013-2017 32