Machine Learning Up to Date #11

A PPO-trained RL agent showing Bowser who’s boss [Source]

Here's ML UTD #11 from the LifeWithData blog! We help you separate the signal from the noise in today's hectic front lines of software engineering and machine learning.

LifeWithData strives to deliver curated machine learning & software engineering updates that point the reader to key developments without superfluous details. This enables frequent, concise updates across the industry without information overload.



Reinforcement Learning (PPO) vs Super Mario

A PPO-trained RL agent showing Bowser who’s boss [Source]
Viet Nguyen on Github, who previously published RL agents trained on the A3C algorithm, has done the same for a PPO algorithm.
... keep reading

Deepmind RL Unplugged

An overview of some of the tasks that DeepMind’s RL Unplugged supports [Source]
Deepmind has released RL Unplugged, which is a suite of benchmarks for offline reinforcement learning tasks. Similar to OpenAI Gym, The library gathers together several common RL tasks and provides a unified API on top of them. Below are additional considerations Deepmind took in selecting the supported tasks: High dimensional action spaces, for example the locomotion humanoid domains, we have 56 dimensional actions. High dimensional observations. Partial observability, observations have egocentric vision. Difficulty of exploration, using states of the art algorithms and imitation to generate data for difficult environments. Real world challenges.
... keep reading

Google AI Tops Nearest Neighbors Benchmark

A visual depiction of neural-network-based embeddings for MIPS [Source]
Google AI released a new algorithm for approximate nearest neighbors (ANN), which is superior to most on the glove-100 benchmark on [ann-benchmarks.com](https://ann-benchmarks.com). While some algorithms use trees or graphs to make the neighbor search more efficient, this algorithm uses the codebook algorithm, which compresses the data set into a fixed codebook size for future searches.
... keep reading

Marriage & Divorce -- By the Numbers

Divorces per 1,000 residents by country [Source]
The good folks over at Our World in Data published a nice report of trends in marriage and divorce over the past ~70 years. Here’s the executive summary below; head to the article for beautiful plots! Marriages are becoming less common: in most countries the share of people getting married has fallen in recent decades. However, this is not true across all countries. Across most countries, people are marrying later in life. Cohabitation – couples living together who are not married – is becoming increasingly common. Single parenting is common and has increased in recent decades across the world. The Netherlands was the first country to legally recognise marriage for same-sex couples in 2000. Since then at least 30 countries have followed suit. There has been a general upward trend in divorce rates globally since the 1970s. But this pattern varies significantly country-to-country. Divorce rates are lower in younger cohorts. In rich countries with available data the average length of marriage before divorce has been relatively stable in recent decades, and in some cases it has even increased.
... keep reading

A Tour of End-to-end Machine Learning Systems

A diagram of the typical machine learning engineering flow [Source]
Ian Hellstrom in the Databaseline blog wrote up a wonderful summary of end-to-end machine learning platforms. To those not part of an industry group trying to maintain ML systems in production, ML Engineering and ML Ops are indeed the hottest data-related job nowadays. Learning to bridge the gap between new research and product delivery will make you a very special employee (or business owner!).
... keep reading

Geometric Deep Learning — Going Beyond Euclidean

Some visualized examples of non-euclidean data [Source]
Some Machine Learning greats teamed up for a recent paper, which provides an overview of the nascent area of "geometric deep learning”, which aims to extend the wide success of deep neural networks to non-euclidean data. Non-euclidean data includes social networks, sensor networks, biological networks, and mesh surfaces in computer graphics. Therefore, innovation in this space will yield great results for the world.
... keep reading