Machine Learning Blueprint Newsletter, Edition 12, 12/31/17

This newsletter is written and curated by Mike Tamir and Mike Mansour. 

December 31, 2017

Hi all,

Please forward to your friends and help us grow our 12k person audience as we start the new year!

Spotlight Articles

Google releases the (very) long awaited CoLaboratory (google docs for jupyter notebooks) where users can use google’s virtual machines to carry out machine learning tasks and make models without worrying about serving compute time. Best part is that it appears to be free (for now).

Human learning takes advantage of existing skills and builds new capabilities by composing and combining simpler ones. Inspired by this observation, the Einstein Labs team proposes a hierarchical RL approach which can reuse previously learned skills alongside and as subcomponents of new skills. It achieves this by discovering without human input the underlying hierarchical relations between skills. As the RL model executes a task, it breaks the task into smaller parts and eventually basic actions using a hierarchy of “policy neural networks” that predict actions, operating at different levels of abstraction.

Machine Learning Blueprint's Take

Learning Machine Learning

A guide to maximizing interpretability in your machine learning models. It focuses less on the math, and more on the process and choices in models.

Machine Learning Blueprint's Take

A gentle introduction to GANs focused some novel application spaces like oncology and pharmacology.

See also a nice reminder about the multiple p-value comparison problem, and a good reason to read up on Bonferroni correction: fMRI Gets Slap in the Face with a Dead Fish

Machine Learning News

Current RL libraries offer parallelism at the level of the entire RL loop, making existing implementations difficult to extend, combine, and reuse. Ray RLLib is a new library aimed at solving this problem by encapsulating parallelism and resource requirements within individual components. The library implements a wide range of state-of-the-art algorithms by composing and reusing a handful of standard components.

The creators of the Stanford NLP project (and NLTK) release a new toolset for analyzing the evolution of language through time by looking a changes word-embeddings. If not applicable to one’s own projects, the published findings remain interesting.

After raising much stir over their comments at NIPS2017, Ali Rahimi & Ben Recht follow up with on their comments. They clarify their definition of “rigor”, backup claims about gradient descent and revisit a statement on batch normalization.

Interesting Research

Google combines Tacotron and Wavenet to synthesize speech from text in a sequence-to-sequence network that is able to capture the finer subtleties in human speech pattern.

Machine Learning Blueprint's Take