Google unveiled a lot of cool stuff @ Google I/O:

• Google Cloud TPUs (Tensor Processing Units) that are capable of 180 teraflops each, and pods with 64 TPUs capable of 11.5 petaflops are now available on the Google Compute Engine. You can read an in-depth article on it here: Under The Hood Of Google’s TPU2 Machine Learning Clusters. Google is also giving away from free compute time on these TPUs via the The TensorFlow Research Cloud.
• Google’s TensorFlow Lite brings machine learning to (cheap) Android devices
• Google also released a dataset with 50 million drawings (doodles) made by the public: Quick, Draw! The Data

News (non-Google!):

• OpenAI released Roboschool, an open-source software for robot simulation that works with the OpenAI gym. They also announced their progress on training robots which can imitate humans from single demonstration.
• Cardiogram trained a deep neural network on the Apple Watch’s heart rate readings and was able to detect atrial fibrillation — the most common heart arrhythmia — with higher accuracy than previously validated methods. They managed to obtain an AUC of 0.97, enabling us to detect atrial fibrillation with 98.04% sensitivity and 90.2% specificity. Read more here: Applying Artificial Intelligence in Medicine

Tutorials and projects:

• Although recurrent neural networks such as LSTMs and GRUs are quite popular for text, CNNs still win in a lot of use-cases because of faster training and parameter sharing. This article provides an overview of Convolutional Neural Networks for Text.
• Implementing CycleGAN in TensorFlow.
• Picasso: A free open-source visualizer for CNNs and its accompanying research paper on arXiV: A Neural Network Visualizer
• Even today most Chatbots are based on keywords and hard-coded responses. Here is a project on GitHub which is trying to make them be more deep thinkers: DeepChatModels: Conversation models in TensorFlow.

Research:

• [1705.04862v2] Efficient Parallel Methods for Deep Reinforcement Learning: This paper reduces the training time for state-of-the-art atari performance to 8 hours on a single GPU. The source code is available here.
• [1705.06640v1] Automated Whitebox Testing of Deep Learning Systems: This paper introduces unit tests for neural networks! The goal is to automate the testing of DL models (specially important when being used in safety-critical situations such as self-driving cars).
• [1705.06366v1] Automatic Goal Generation for Reinforcement Learning Agents: Its a common problem in RL that the learning system doesn't learn anything useful because the goal is too difficult to get to and it doesn't have any positive reinforcement. This paper aims to have a generator network which generates smaller goals which are of desired level of difficulty for the RL agent, thereby providing a curriculum.

So much cool stuff in one week! Hope you enjoyed!