Issue #11: Dark Data, AI Talent, and Reinforcement Learning

female engineer controlling flight simulator

Subscribe • Previous Issues

This edition has 817 words which will take you about 4 minutes to read.

“When someone’s always watching, we lose our sense of self.”  – Patricia Williams

Data Exchange podcast

Image from Deedee86 from Pixabay

[Image: Deedee86 from Pixabay]

Machine Learning tools and infrastructure

  • Ten Questions on AI Risk  A new checklist from the Future of Privacy Forum.
  • Attacking Deep Reinforcement Learning  RL has been applied to autonomous driving and automated trading, domains where it would be hard for an attacker to directly modify the subject’s policy input. In the self-driving car example, an adversary can impact a camera’s image, but only in a physically realistic fashion (an adversary cannot add noise to arbitrary pixels or make a building disappear). This UC Berkeley research project involving simulated robotics games, introduces adversarial policies that reliably beat their victim.
  • Unsupervised Translation of Programming Languages   A new paper from Facebook AI, shows how unsupervised machine translation can be applied to source code to create a transcompiler. This could result in tools that can automatically migrate existing codebases to modern or more efficient programming languages. 
  • Demystifying AI Infrastructure   This short video describes a landscape map that brings greater clarity to the AI ecosystem.
  • Enterprise Applications of Reinforcement Learning   This new seven minute video covers applications of RL to recommenders and simulation software.


Virtual Conferences


Work and Hiring

[Image: AWeith / CC BY-SA]


  • Dark Data: Why What You Don’t Know Matters   Add this to the list of books to give to your CxO. Statistician David Hand develops a taxonomy for “dark data”, his shorthand for different types of missing data. This book is aimed at a non-technical audience and is filled with examples of what can happen when people fail to acknowledge that they likely have incomplete data. A much cited quote about COVID-19 from this week (“if we stop testing right now we’d have very few cases”) makes the case for sharing this book widely.
  • One Simple Chart: News Consumption in the US  A group of researchers combed through Google Scholar and found an explosion of research on online sources of fake news and misinformation. The goal of the study was to put fake news into context, and the authors concluded that too much attention is placed on fake news. They found that (1) news consumption is heavily outweighed by other forms of media consumption, (2) Americans consume more TV news than online news, (3) fake news is a small part of the overall media diet.  Note that this study took place from January 2016 to December 2018, long before COVID-19 and shelter-in-place may have affected media consumption habits. A March/2020 analysis hinted that news has become America’s biggest pastime.
  • Penrose from CMU – from mathematical notation to beautiful diagrams  This new project lets you translate abstract statements written in math-like notation into one or more visual representations.
  • H.E.R’s Full Performance On Graduate Together 2020   My favorite performance at the recent #GraduateTogether virtual celebration was H.E.R. on piano singing “Sometimes”. Here’s a good interview with H.E.R. recorded in a Filipino restaurant in NYC.