Keynote

Live from Mountain View, CA! Join the TensorFlow team as they kick off the 2018 TensorFlow Dev Summit! The TensorFlow Dev Summit brings together a diverse mix of machine learning users from around the world for a full day of highly technical talks, demos, and conversations with the TensorFlow team and community.

tf.data: Fast, flexible, and easy-to-use input pipelines

Derek Murray discusses tf.data, the recommended API for building input pipelines in TensorFlow. In this talk, he introduces the tf.data library, and presents some recent developments that have improved its performance, flexibility, and ease-of-use.

Eager Execution

Alex Passos discusses Eager Execution, which provides a simpler, more intuitive interface to TensorFlow. This interface is more "Pythonic" and does away with the distinction between constructing and executing computational graphs.

Machine Learning in JavaScript

Nikhil Thorat and Daniel Smilkov discuss TensorFlow.js, which is TensorFlow’s new machine learning framework for JavaScript developers. It supports building ML models in JavaScript, and training and deploying them in browser for execution using WebGL. They focus on tensorflow.js- Core which is a rebranding of deeplearn.js, and a new high-level layers API to make it easier to develop ML models in JavaScript. Initial set of converters for converting saved TF models into JS is also being released.

Training Performance: A user’s guide to converge faster

Brennan Saeta walks through how to optimize training speed of your models on modern accelerators (GPUs and TPUs). Learn about how to interpret profiling tools and techniques to parallelize and overlap work to converge faster!

The Practitioner's Guide with TF High Level APIs

Mustafa Ispir discusses high level APIs which let ML practitioners do many more modeling experiments with only a few lines of code. Some highlights are Head API, Transfer Learning, and Gradient Boosted Trees.

Distributed TensorFlow

Igor Saprykin offers a way to train models on one machine and multiple GPUs and introduces an API that is foundational for supporting other configurations in the future.

Debugging TensorFlow with TensorBoard plugins

Watch this demo of the TensorFlow Debugger, an interactive web GUI for controlling the execution of TensorFlow models, setting breakpoints, stepping through graph nodes, watching tensors flow in real-time, and pinpointing problems down to the tiniest NaN. This tool now comes included with TensorBoard via its open plugin API.

TensorFlow Lite

Sarah Sirajuddin and Andrew Selle discuss TensorFlow Lite, which was announced in developer preview in November 2017. It is a lightweight library that includes associated tools for on-device machine learning on Android, iOS, or smaller devices.

Searching Over Ideas

Getting the most out of Machine Learning models requires careful tuning of many knobs. In this short talk, Vijay Vasudevan discusses the opportunity for turning more of our ML ideas into such searchable components, and using AutoML techniques to discover better models.

Reconstructing Fusion Plasmas

Ian Langmore reconstructs plasma temperature, density, and B-field from measurements in partnership with tae.com. This is a Bayesian inverse problem, making use of TensorFlow's distribution and tensorflow_probability libraries.

Nucleus: TensorFlow toolkit for Genomics

Cory McLean announces the launch of Nucleus, a library of Python code for reading, writing, and filtering common genomics file formats for conversion to TensorFlow examples. Cory briefly describes its role in creating DeepVariant, an open-source TensorFlow CNN-based program for genome variant discovery that substantially improves upon prior methods.

Open Source Collaboration

Edd Wilder-James announces a new set of mailing lists to help communication and coordination, the expansion of the SIG program (Build, Swift, JavaScript, TensorBoard and Rust), and previews the public RFC process (to come in April/May 2018).

Swift for TensorFlow - TFiwS

Chris Lattner and Richard Wei unveil Swift for TensorFlow (TFiwS) as an early stage open source project. It has many design advantages, and will be released with technical whitepaper, code, and an open design approach in April of this year. TFiwS means TensorFlow integrated with Swift but we are officially calling it Swift for TensorFlow. Stay tuned for more announcements from the TensorFlow team regarding the support of Swift!

TensorFlow Hub

Andrew Gasparovic and Jeremiah Harmsen dicuss TF Hub, a new library built to foster the publication, discovery, and consumption of reusable parts of machine learning models. The main part of the launch is a repository of modules, which are self-contained pieces of TensorFlow graphs that can be reused across different tasks. Modules contain variables that have been pre-trained for a task using a large dataset. By reusing a module on a related or similar task, a user can train a model with a smaller dataset, improve generalization, or simply speed up training.

TensorFlow Extended (TFX)

Clemens Mewald and Raz Mathias present TFX, which is an end-to-end ML platform built around TensorFlow and first introduced to the public in a 2017 KDD paper. While TF.Transform and TF.Serving are already open sourced, Clemens introduces a new component, TensorFlow Model Analysis (TFMA), and give an end-to-end demo of how those tools fit together. They also announce plans about releasing more of TFX.

Applied AI at The Coca-Cola Company

In 2016, Coca-Cola updated its core loyalty marketing program to a mobile-first web platform. The program requires consumers to input 14- character proof-of-purchase codes for entry into promotions. Consumer engagement would have been badly impacted if users had been required to thumb-enter these codes into web forms on their mobile devices. To solve this problem, Coca-Cola built a custom Optical Character Recognition (OCR) capability using Convolutional Neural Networks and TensorFlow. This TensorFlow CNN is hosted online for web-based promotions and installed natively as a component of the Coca-Cola iPhone and Android apps. Patrick Brandt of Coca-Cola North America will share his design for Coke’s mobile proof-of-purchase platform and discuss some of the challenges that his team overcame to deliver a fast, highly-accurate custom OCR using TensorFlow.

Real-World Robot Learning

Alex Irpan discusses real-world robot learning. In the past, research has shown that with enough real world robot data, we can teach a real robot how to pick up objects. Alex shows how the team used physics simulators and other ML techniques to reduce the amount of real world data required.

Project Magenta

Magenta explores the role of ML in the process of creating art and music. This involves developing new deep learning and reinforcement learning algorithms for generating songs, images, drawings, and other materials. It's also an exploration in building smart tools and interfaces that allow artists to extend (not replace!) their processes.