Google I/O Extended 2019 - GDG Seattle

Transcript

1. Machine Learning at Google I/O 2019 Niveditha (Nivii) Kalavakonda, Ph.D. Student University of Washington, Seattle @nkalavak
2. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Overview 1. Interesting ML Projects 2. Google AI Impact Challenge 3. TensorFlow 2.0 4. Cutting Edge TensorFlow 5. Human-Centered AI 6. Federated Learning 7. Other fun I/O stuff!
3. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers My First I/O!
4. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Interesting ML Projects @ I/O Google Lens Credit:https://9to5google.com/2019/05/21/google-lens-revamp/
5. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Interesting ML Projects @ I/O Google Lens Credit:https://www.gsmarena.com/google_lens_update_brings_new_design_and_dinning_mode_feature-news-37170.php
6. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Interesting ML Projects @ I/O Speech Recognition • Live Transcribe • Live Caption and Live Relay (runs on-device inference) • Project Euphonia
7. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Interesting ML Projects @ I/O Google AI Research • Lung Cancer Prediction and Diabetic Retinopathy • Computer Vision and Robotics (Grasping, reinforcement learning to pour fluids, etc) • BERT: State-of-the-Art Pre-training for NLP • Per-Example Routing for Machine Translation • AutoML - Neural Architecture Search ◦ Learning optimization update rules, non-linearity for activation functions and other updates • Cloud TPU v3 - Runs same code as v2 with water cooling. 420 teraflops, 128 GB HBM (To train & infer) Link to Talk
8. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Google AI Impact Challenge • Part of Google’s AI for Social Good program to fund projects that address social, humanitarian and environmental problems • 20 winners chosen out of 2602 applications from 119 countries! • Read about the projects: https://ai.google/social-good/impact-challenge/ Highlights: Talking Points (USA): • Messaging app between parents and students to improve conditions in low-income communities • Provides personalized, real-time coaching content to parents and teachers • Continuously improves the translation quality on the TalkingPoints family engagement platform Wadhwani AI (IND): • Farmers struggle with avoidable pest damage that can wipe out 50% annual crop yield • AI models will classify and count the pests from regular photos of pest traps • Data collected will be used to provide millions of farmers with advice to reduce pesticide usage and improve yield La Fondation Médecins Sans Frontières(FR) • Medical staff need to prescribe the right antibiotic for bacterial infections • Antibiograms help determine resistance but in low-resource areas, skill to use it is lacking • Using image recognition and a smartphone app, the project will help staff analyze anti-microbial images for providing treatment Link to Talk
9. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers TensorFlow 2.0 • What’s new in TF2.0? ◦ Easier to learn and use ◦ For beginners and experts ◦ Currently in alpha • What’s the appeal? ◦ tf.Keras is the recommended high level API and eager execution as default (Usability) ◦ Declutters the duplicated APIs for consistency (Clarity) ◦ Continues to offer access for those who need low level APIs (Flexibility) • Installation: • To get started, you can follow a free Udacity course or audit on Coursera • To upgrade: ◦ Escape to backwards compatibility module (tf.compat.v1) ◦ Conversion scipt: tf_upgrade_v2 (or) use TF2up ◦ Guides and best-practices in TF2.0 documentation !pip install --pre -U tensorflow Link to Talk
10. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers TensorFlow 2.0 • TF2.0 Roadmap: ◦ Link: https://www.tensorflow.org/community/roadmap Credit:https://medium.com/tensorflow/whats-coming-in-tensorflow-2-0-d3663832e9b8 Link to Talk
11. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Cutting Edge TensorFlow • Keras Tuner ◦ Hypetuning framework for humans ◦ Steps: Wrap model in a function, define hyper-parameters and replace static value with hyper-parameters ◦ Will also include a model dashboard to track tuning! ◦ Link(Early access): https://g.co/research/kerastunereap • Probabilistic Programming ◦ TensorFlow Probability: Toolbox for Probabilistic modeling ◦ Has tools to build models and run inference from those models ◦ Link: https://www.tensorflow.org/probability • Ranking with TF-Ranking ◦ TF-Ranking: A Scalable learning-to-rank library (using deep learning approaches) ◦ Support many loss functions and metrics for optimization ◦ Compatible with TensorBoard and TensorFlow Serving ◦ Link: https://ai.googleblog.com/2018/12/tf-ranking-scalable-tensorflow-library.html Link to Talk
12. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Cutting Edge TensorFlow • TensorFlow Graphics ◦ Link: https://github.com/tensorflow/graphics/ Credit:https://medium.com/tensorflow/introducing-tensorflow-graphics-computer-graphics-meets-deep-learning-c8e3877b7668 Link to Talk
13. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Human-Centered AI At Google: People + AI Research (PAIR) team • People + AI Guidebook ◦ Link: https://pair.withgoogle.com/ ◦ Works with the AI principles and TensorFlow/ML Kit/Other Frameworks and Tools • Guidebook covers six aspects: ◦ User Needs + Defining Success ◦ Data Collection + Evaluation ◦ Explainability + Trust ◦ Errors + Failing Gracefully ◦ Mental Models ◦ Feedback + Control • Tools: ◦ Facets - To understand your data ◦ What-if? - Analyze an ML model ◦ Distill.pub - Projects on explainability ◦ Other PAIR projects Link to Talk
14. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Federated Learning What is Federated Learning? Machine Learning on decentralized data Why Federated Learning? • Smarter models • Lower latency • Less power consumption • Privacy-by-default TensorFlow Federated: • Experiment with federated technologies in simulation • Offers Federated Learning API and Federated Core API • Link: https://www.tensorflow.org/federated • Comics on Federated Learning: https://federated.withgoogle.com/ Link to Talk
15. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Other Talks of Interest Technical Talks • Swift for TensorFlow • TF-Agents: A Flexible Reinforcement Learning Library for TensorFlow • Machine Learning for Game Developers • Machine Learning Zero to Hero • Cloud TPU Pods: AI Supercomputing for Large Machine Learning Problems • Deep Learning to Solve Challenging Problem • Machine Learning Fairness: Lessons Learned Non-technical Talks • A Fireside Chat with Turing Award Winner Geoffrey Hinton, Pioneer of Deep Learning • A Fireside Chat with X’s Captain of Moonshots - Astro Teller • Artificial Intelligence: From Social Good to Ambient Intelligence • Making Art with Artificial Intelligence: Artists in Conversation
16. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers ML Codelabs and Sandbox Link: https://codelabs.developers.google.com/io2019/
17. @nkalavak | #io19 | #io19Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers ML Codelabs and Sandbox
18. Huge thanks to Paige Bailey and AnitaB.org!