Monthly MLAI Meetup

Kristian Holsheimer on Keras-gym: Plug-n-play Reinforcement Learning in Python

Reinforcement learning (RL) is cool.. but it's hard. One of the reasons why it's hard is that there aren't many packages that will allow you to only implement the logic you care about. You usually end up writing many lines of the same boilerplate code. Compare this to supervised learning, where training a model is as easy as "import foo; foo.fit(X, y)". The keras-gym package is designed to close that gap for RL. In this talk you'll see how easy it can be to train an RL agent yourself.

Bio: Kris is an applied machine learning scientist at Microsoft, mostly working on protecting Windows users against malicious or otherwise unwanted software. The reason for his interest in reinforcement learning is that touches on many different areas of machine learning. Kris received his PhD from the University of Amsterdam on the topic of string theory.

David Rawlinson on Learning distant cause and effect with only local and immediate credit assignment

Deep backprop is a near universal assumption of modern ANNs. But there's little evidence that wet brains have any signalling mechanism that could approximate it. This motivates our search for an alternative using only local credit assignment to train model parameters. I'll describe a simple model and show tantalizing results on a few benchmarks.

Bio: 20+ years FT R&D in AI/ML - mostly applied, some academic (NICTA, Melb U.). PhD robotics / vision based navigation @ Monash U. BSc Comp. Sci. & AI (COGS, Sussex, UK). Certified Chainsaw Operator.