Episode 22: Archit Sharma, Stanford, on unsupervised and autonomous reinforcement learning
Generally Intelligent
00:00
Learning a Gradient Signal in a Neural Network
The idea here was to learn an gradient signal across discrete latent variables. The whole pair was that it might introduce a bit of bias because any function of approximation is biased, but it might cut down on the variance. It ended up being the case that the bias is actually too high. Even that learn faster in the beginning ended up like the some optimal performance.
Transcript
Play full episode
