Controlling a unicycle with Policy Gradients

  1. Given that passive dynamic walkers and unicycles are dynamically similar to inverse pendulums, investigations on unicycle dynamics naturally lead to experiments on bipedal dynamics.
  2. Due to favorable trade-offs of effectiveness and model complexity I decided to start modelling unicycle behaviour using Policy Gradient methods.

code: Controlling a unicycle with Vanilla Policy Gradients
blog post: Controlling a unicycle with Policy Gradients

Understanding the free energy principle

  1. The free energy principle is a theory developed by Karl Friston and others to explain how biological systems tend to avoid disorder by limiting themselves to a small number of favorable states.
  2. Until recently, the Free Energy Principle has been a constant source of mockery from neuroscientists who misunderstood it and so I hope that by growing a collection of free-energy motivated reinforcement learning examples on Github we may finally have a constructive discussion between scientists.
  3. Regarding mockery, there is even a satirical Karl Friston account on Twitter.

code: Free Energy Experiments
blog post: Understanding the free energy principle

Is Causal Path Entropy an E=mc^2 for intelligence?

  1. The Causal Entropic Force framework proposes that organisms make decisions which maximise their future optionality.
  2. Alex Wissner-Gross, the main author of this theory thinks that this is the closest thing to an E=mc^2 for intelligence.
  3. This repository is designed to help deconstruct this claim by analysing simple problems which provide insight into the limitations of Causal Entropic Forces.

code: Is Causal Path Entropy an E=mc^2 for intelligence?
blog post: Approximating Causal Path Entropy in Euclidean spaces

Fractals with TensorFlow:

  1. Right now, this repository is partly used to explore the symmetries of Mandelbrot variants.
  2. In the near future, I’ll add a GPU optimised version for investigating quaternion fractals.
  3. I wrote an introductory blog post for this repository.

code: Fractals with TensorFlow
blog post: Fractals with TensorFlow

Deep Rectifiers:

  1. A set of tools for analysing deep fully-connected recitifer networks i.e. networks with ReLU activation.
  2. Specific tools include methods for analysing optimisation, variable disentangling, variable size representation, sparsity, and generalisation.
  3. The ultimate goal is to develop useful mathematical models for deep rectifier networks, a model for supervised learning that is not yet well-understood.

code: deep_rectifiers
blog post: deep rectifier networks: preliminary observations