Machine Learning Blueprint Newsletter, Edition 18, 4/1/18

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

April 1, 2018

Hi all,

Hope you enjoy this week's ML Blueprint. This week is brought to you by Fast Data IO.

Spotlight Articles

A detailed but clearly articulated exploration into using generative neural networks to enhance video game playing reinforcement learning (RL) models. Using a technique called variational autoencoders, they equip the RL agent with a "world model" in advance of the RL process. They take these results a step further, using the world models of the agents to artificially generate "dream training" environments in which the RL agent trains. Their experimental results show that not only does a world model enable the agent to outperform other leading models, but by allowing the RL agent to train in dream environments that are made artificially "noisier" than reality the agents were able to learn even more successfully.

Machine Learning Blueprint's Take

AI has the power to influence human behavior, mainly by controlling the feed of information that we receive. This becomes a feedback loop that can be optimized w.r.t. Achieving a certain state via action, where now humans are the agent in this global reinforcement learning experiment. This means that the controllers of this information can manipulate the users into desired behavioral patterns.

Machine Learning Blueprint's Take

Now witness the real-time power of this fully GPU-armed and operational stream processing engine.

FDIO Engine™ can process Spark Structured Streaming workloads up to 1000x faster than Apache Spark, as it runs natively on NVIDIA GPUs and Apache Arrow.

To put 1000x performance in perspective, FDIO processes 1 Terabyte on a single AWS instance in 35 seconds, where Spark takes over 9 hours.

Sign up for a Test Flight today at fastdata.io

Learning Machine Learning

Square is using this method to evaluate users flagged by some detection model (probably not a NN) to give better reasons for why they are anomalous, but with the requirement of doing it very quickly. They train a GAN to generate synthetic “good” users, and then compare a real “bad” one to the K closest neighboring synthetic “good” users. Next, the features between the average of the K good fake users and the real bad one are compared to identify the one that most likely triggered the anomaly.

Machine Learning Blueprint's Take

This research attempts to mimic the way humans learn to navigate a new environment; the problem is akin to a maze navigation problem. They train a 3-part network to learn how to traverse across the city with Google maps street-view with first a CNN to process and featurize visual inputs, a locale-specific RNN that learns the environment, and lastly an RNN for navigation policies over actions. In the end, they’re able to successfully navigate across several different cities. The model can be made transferable by freezing the 1st and 3rd components and retraining the locale-based RNN.

A fellow machine learning denizen constructed a Jupyter Notebook series to jump start someone trying to get into Deep Learning through Goodfellow’s & Bengilos pretty good open source book.

Machine Learning News

Emmanuel Macron realizes that the US & China are leading on AI, and he is creating a plan that tacts between the two powerhouses. He believes that AI will have profound effects on democracy; to ensure that the effects are positive, both government and private institutions should be transparent about algorithms and opening portals to data. His country will be accepting of AI integration on all fronts except within military applications.

Machine Learning Blueprint's Take

The TF dev team announced a number of new and upcoming features: TensorFlow Hub for sharing modules, tools for better visualization evaluation metrics, enhanced Nvidia GPU & and Intel CPU support, TF in Javascript, and new upcoming libraries for geomics.

If not already familiar with Alibaba beyond being an online retailer, they also have a massive cloud AI strategy that is rivaling territories for both US giants Amazon and Google. Alibaba is slowly catching up in R&D spending, spending $2.6B last year, while Google and Amazon spent ~$15B each. Looking to their cloud strategy, on top of offering off-the-shelf ML API’s, Alibaba supports all major ML frameworks on their hardware, eliminating any lock-in (however, this might be moot with the proliferation of ONNX).

Leveraging facial recognition algorithms, the government is able to identify jaywalking citizens, send them a fine, publicly shame them, and affect their ability to get a job or loan. They’ve already caught 14,000 criminal jaywalkers at a single intersection in Shenzhen.

Machine Learning Blueprint's Take

NIH published their strategic plan for bringing AI into their domain to help push the research boundary. They requested feedback from data science community stakeholders about the plan.

Interesting Research

The authors find a useful way to vectorize series of paths through the Abstract Syntax Tree of code. The vectorized representation can be used to describe certain properties of the code. The authors measure the algorithm’s effectiveness by seeing how well it can predict the method name.