Up to Speed on Deep Learning: September Update, Part 1

Continuing our series of deep learning updates, we pulled together some of the awesome resources that have emerged since our last post on August 31st. In case you missed it, here are our past updates: August part 2, August part 1, July part 2, July part 1, June, and the original set of 20+ resources we outlined in April. As always, this list is not comprehensive, so let us know if there’s something we should add, or if you’re interested in discussing this area further.

One Hundred Year Study on Artificial Intelligence (AI100) by the AI100 Standing Committee (full PDF here). The One Hundred Year Study on Artificial Intelligence, or AI100, is a 100-year effort to study and anticipate how the effects of artificial intelligence will ripple through every aspect of how people work, live and play. Researchers explain opportunities to apply artificial intelligence across many domains, such as transportation and healthcare, and the ways in which it has and will affect our lives. They also share a framework around AI policy. The group plans to assess the state of AI every five years.

Infrastructure for Deep Learning by the OpenAI team. In this post, we’ll share how deep learning research usually proceeds, describe the infrastructure choices we’ve made to support it, and open-source kubernetes-ec2-autoscaler, a batch-optimized scaling manager for Kubernetes.

TF-Slim: A high level library to define complex models in TensorFlow by Nathan Silberman and Sergio Guadarrama of Google Research. TF-Slim is a lightweight package for defining, training and evaluating models in TensorFlow. The Google Research team announces the latest release of TF-Slim, which includes many types of layers, loss functions, evaluation metrics, and handy routines for training and evaluating models.

PaddlePaddle by Baidu (GitHub repo here). Open and Easy-to-Use Deep Learning Platform for Enterprise and Research. PaddlePaddle (PArallel Distributed Deep LEarning) is an easy-to-use, efficient, flexible and scalable deep learning platform, which is originally developed by Baidu scientists and engineers for the purpose of applying deep learning to many products at Baidu.

Deep Learning in a Nutshell: Reinforcement Learning by Tim Dettmers of NVIDIA. A broad overview of reinforcement learning including the major concepts and areas in which the technique shines. The 4th installment in Tim’s excellent series of articles that explain deep learning core concepts.

WaveNet: A Generative Model for Raw Audio by DeepMind (original paper here). The team presents a model called WaveNet, which is able to generate speech which mimics any human voice and which sounds more natural than the best existing Text-to-Speech systems, reducing the gap with human performance by over 50%. Also, the same network can synthesize other audio signals such as music, and present some striking samples of automatically generated piano pieces.

Attention and Augmented Recurrent Neural Networks by Chris Olah and Shan Carter. Researchers at Google Brain explain four ways to augment Recurrent Neural Networks with new and more sophisticated properties, as well as the concept of attention that is an underlying principle of each.

Why does deep and cheap learning work so well? by Henry Lin and Max Tegmark. Researchers from Harvard and MIT articulate that the reason behind the efficacy of neural networks can be explained via the laws of physics versus math alone. We show how the success of deep learning depends not only on mathematics but also on physics (summary article here).

How a Japanese cucumber farmer is using deep learning and TensorFlow by Kaz Sato of Google Cloud Platform. The article explains Makoto Koike’s — a Japanese embedded system designer — end-to-end process of building a cucumber sorter that he implemented on his parents’ cucumber farm in Japan. Koike leverages TensorFlow to implement his deep learning based classifier.

Playing for Data: Ground Truth from Computer Games by Richter et al. Researchers at TU Darmstadt and Intel Labs present an approach to creating detailed label maps from the Grand Theft Auto video game due to the game’s high degree of realism. This increases accuracy of self-driving algorithms that are traditionally trained on real-world data sets.


By Isaac Madan. Isaac is an investor at Venrock (email). If you’re interested in deep learning or working on something in this area, we’d love to hear from you.

Subscribe to our email newsletter here. Requests for Startups is a newsletter of entrepreneurial ideas & perspectives by investors, operators, and influencers.