The Tensorflow Developer Summit took place on March 30, 2018. It featured dozens of tutorial sessions and several important announcements, including the release and re-branding of Tensorflow.js, a JavaScript port of Tensorflow, the release of Tensorflow Hub, a service to share Tensorflow model code, and Tensorflow support for SWIFT, among many other things.

DeepMind announced the opening of the first research lab in continental Europe, DeepMind Paris, which will be led by Remi Munos - one of DeepMind’s principal research scientists and author of 150 research papers

The new service allows you to use high-quality text-to-speech synthesis that produces natural sounding speech for your own apps. The pricing is $16.00 USD for 1 million characters.

Can agents learn inside their own “dreams”? In this research, an agent is trained inside a learned environment model, a hallucination. The learned policy is then transferred back into the real environment and still performs surprisingly well. Read the paper here.

A free series of blog posts about Deep Reinforcement Learning, where you can learn about the main algorithms and how to implement them in Tensorflow.

This Youtube channel contains all recordings from the Tensorflow Developer Summit that took place last week. An excellent resource to learn about new Tensorflow features and use cases.

Ever wondered what learning rate to use? 3e-4? Fixed? Cyclic? Fear no longer because the learning rate mystery has finally been solved: Use an aesthetically pleasing learning rate! This will make you happy regardless of what the training result is.

A script to automatically “block” people in images (like the Black Mirror episode White Christmas) using Mask R-CNN pre-trained on the MS COCO dataset.

Matchbox enables deep learning researchers to write PyTorch code at the level of individual examples, then run it efficiently on minibatches.

RL algorithms struggle to solve simple tasks when enough information is concealed from the sensors of the agent, a property called “partial observability”. The authors develop a model, the Memory, RL, and Inference Network (MERLIN), in which memory formation is guided by a process of predictive modeling. MERLIN facilitates the solution of tasks in 3D virtual reality environments for which partial observability is severe and memories must be maintained over long durations. The model demonstrates a single learning agent architecture that can solve canonical behavioral tasks in psychology and neurobiology without strong simplifying assumptions about the dimensionality of sensory input or the duration of experiences.

The authors explore building generative neural network models of popular reinforcement learning environments. The world model can be trained in an unsupervised manner to learn a compressed spatial and temporal representation of the environment. By using features extracted from the world model as inputs to an agent, one can train a compact policy that can solve the required task. The agent can be trained entirely inside of its own hallucinated world model, and transfer this policy back into the actual environment. 

A neural model for representing snippets of code as continuous distributed vectors. The main idea is to represent code as a collection of paths in its abstract syntax tree, and aggregate these paths, in a smart and scalable way, into a single fixed-length code vector, which can be used to predict semantic properties of the snippet. The authors demonstrate the effectiveness of the approach by using it to predict a method’s name from the vector representation of its body.