exploring reinforcement learning with SNNs on SpiNNaker
Reinforcement learning is a well-established mechanism in biology, whereby interacting with your environment leads to learning based on perceived rewards and punishments, for example, through dopamine signalling. These concepts have been widely explored in machine learning to help train deep neural networks to play games such as chess and Go.
The main goal of this project is to explore the training of a brain-like spiking neural network (SNN) to play an Atari game such as Breakout or Pong. The SpiNNaker neuromorphic platform will be used to facilitate massively-parallel online learning.
I am currently working on this as part of my third-year project at the University of Manchester. More information will be available on my blog throughout the year.
- Research a range of neuronal dynamics, such as leaky integrate-and-fire models and spike-time dependency plasticity
- Explore multi-agent reinforcement learning for both cooperative and competitive (zero-sum game) scenarios
- Execute Atari games directly on the SpiNNaker machine
- Implement (in Python and C) and compare the results between different reinforcement learning algorithms, as well as comparing it to other neural network types