Import AI: Issue 1: GANs, ML bias, and a neural net Benjamin Franklin

by Jack Clark

Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.

Adversarial training / generative adversarial networks: “the most interesting idea in the last 10 years in ML” says Yann Lecun, a jazz aficionado who has a day job as Facebook’s Director of AI Research. One problem with GANs is that they are quite unstable and choosing the right settings is currently mostly an act of intuition, kind of like convolutional networks were a decade ago. Onward!

“It’s not my fault my data contains bias” is the new “the dog ate my homework”. Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings? research from Boston University & Microsoft Research.

Machine learning is not, by default, fair or just in any meaningful way,” says Stephen Merity.

Now we know Google has used reinforcement learning to reduce the power consumption of its data centers it’s reasonable to wonder how else RL can and will be applied. Answers so far include robotics, wildfire suppression, healthcare, and more. RL will eventually be used to simulate (and run) complex multi-agent environments, like the power grid. Electric Power Market Modeling With Multi-Agent Reinforcement Learning gives some good clues.

Layer normalization: new technique substantially reduces training time of recurrent nets. On one question-answering task it “trains faster but converges to a better validation result”. Which sounds suspiciously like ‘having cake and eating it too’. From University of Toronto & Google/UofT’s Geoff Hinton. Try it for yourself via this TensorFlow implementation.

The Machine Intelligence Research Institute has accomplished a lot in the last year as it grapples with the paradoxes of controlling superintelligence. Its greatest achievement, though, is the invention of the term “Vingean Reflection”.

Riddle me this: ““Joan made sure to thank Susan for all the help she had given. Who had given the help? Answer 0: Joan or Answer 1: Susan””. You probably got this right. A computer would probably get this wrong, according to the latest results of the Winograd Schema Challenge.

You should follow @smilevector on twitter. It’s an experiment from Tom White that uses modern AI techniques to manipulate faces. It certainly cheered up Benjamin Franklin! (Though I’m less sure about Obama.

Thanks for reading. If you have suggestions, comments or other thoughts you can reach me at or tweet at me @jackclarksf