TensorFlow Dev Summit 2018 Niveditha (Nivii) Kalavakonda, Ph. D. Student University of Washington @nkalavak 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Topics of Interest 1. TensorFlow for Healthcare 2. tf.data 3. TensorFlow.js 4. Searching over Ideas 5. Nucleus: TensorFlow Toolkit for Genomics 6. Real-World Robot Learning 7. Project Magenta 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers TensorFlow for Healthcare 1. Ophthalmology - Diabetic Retinopathy a. Automated Retinal Disease Assessment (Talk from 2017 by Dr. Lily Peng: https://goo.gl/XAYLGf) b. Predict things doctors can’t - Age, Gender, Haemoglobin level and other factors (Link: https://goo.gl/XAvz9M) 2. Predictive tasks for healthcare a. Predict health given patient’s electronic medical record data b. Most likely length of stay, diagnosis, tests,etc., c. Link: https://arxiv.org/abs/1801.07860 (At-risk in next 24 hours) Weighted Kappa Opthamologists 0.80-0.84 Algorithm 0.84 Retinal Specialists 0.82-0.91 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers tf.data Recommended API for building input pipelines in TensorFlow ● Pipeline should be Fast, Flexible (diverse data and use cases) and easy to use ● CNN benchmarks : >13,000 images/second with tf.data ○ Throughput doubled over last 8 months (Link: https://goo.gl/if13Ld) ● Input pipeline performance guide (Link: https://goo.gl/fo4UvG) ● GPU prefetching optimization, Num_parallel_reads, fused transformations ● Backend API access - Can now build Dataset Kernel plugins using C++ ● Eager Execution - Treat datasets as iterable objects Example code: github.com/tensorflow/bechmarks and github.com/tensorflow/models Watch Talk here and Brennan Saeta’s Training Performance talk here 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers TensorFlow.js ● Open-source library to define, train and run machine learning models entirely in the browser ● Uses Javascript and high-level Layers API ● Three workflows you can consider: ○ Import an existing, pre-trained model for inference ○ Re-train an imported model ○ Author models directly in browser ● Examples ○ TensorFlow Playground (https://goo.gl/HS97Ta) ○ Pacman (https://goo.gl/Z4BTVi) ○ Emoji Scavenger Hunt (https://emojiscavengerhunt.withgoogle.com/) Links: https://js.tensorflow.org/ https://github.com/tensorflow/tfjs Mailing list: goo.gl/drqpT5 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Searching over Ideas Designing search spaces over solutions to ML problems and using automated exploration How to decide learning rate, dropout rate and other variable parameters (researcher intuition, trial and error, hyperparameter optimization) Which sub-architecture to pick, activation functions to use? Search space for Convolutional cells- Select inputs, operations, repeat and concatenate. ● Many combinations to generate there cells 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Searching over Ideas ● Program generator generates samples from search space ● Train and evaluate on task ● Optionally uses feedback - Reinforcement learning, evolutionary algorithms, or random search ● Can also learn optimization update rule Can optimize on accuracy and inference speed Link: https://goo.gl/etfceM 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Nucleus: TensorFlow Toolkit for Genomics ● Google Brain Genomics team ● Library of Python code for reading, writing, filtering genomics file formats for conversion to TensorFlow Examples ● DeepVariant ○ Open-source TensorFlow CNN-based program for genome variant discovery ○ Link: github.com/google/deepvariant ● Interest in TensorFlow for Genomics: ○ Open Source - Libraries for working with genomics in TF ○ Example genomics tools using TF ● Nucleus (C++/Python library) : https://github.com/google/nucleus ● Cloud integration with Variant Transforms: github.com/googlegenomics/gcp-variant-transforms 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Real-World Robot Learning ● Google Brain Robotics Team ● Simulation and domain adaptation in real-world robot learning ● Task: ○ Robot arm to grasp an objects inside a bin ● Methods ○ The Arm Farm ○ Using Simulation ○ Sim-to-Real Transfer ■ Randomized simulations (Texture, Colors, Lighting) ■ Object Models (Geometry) ■ Domain adaptation (feature-level and pixel-level) 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Real-World Robot Learning Feature-level Domain adaptation ● Similarity Loss is introduced at intermediate layer ● Domain-Adversarial Neural Networks Pixel-level Domain adaptation ● Transform images at pixel-level to look more realistic Combine both methods - Grasp GAN Matches real-only performance with 50x fewer real samples! Link: https://goo.gl/G1HSws 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Project Magenta Explores the role of Machine Learning in the process of creating art and music ● Performance RNN ● Variational Autoencoders ○ NSynthSuper (Moves one input to another instead of recomposing) ■ Link: https://github.com/googlecreativelab/o pen-nsynth-super ○ Music VAE - Gradual blending of 2 different melodies Link: https://magenta.tensorflow.org 

@nkalavak | #TFDevSummit | #Extended | @GDGSeattle | @WTMSeattle | #WomenTechMakers Women Techmakers Travel Grant recipients: 2017 - 5 grants awarded 2018 - 9 grants awarded Attendees from different graduate schools and postdoctoral positions Join us! https://www.womentechmakers.com/