October 26, 2020

Biological Reinforcement Learning via Predictive Spacetime Encoding

Recent advances in reinforcement learning (RL) have successfully addressed several challenges, such as performance, scalability, or sample efficiency associated with the use of this technology. Although RL algorithms bear relevance to psychology and neuroscience in a broader context, they lack biological plausibility. Motivated by recent neural findings demonstrating the capacity of the hippocampus and prefrontal cortex to gather space and time information from the environment, this study presents a novel RL model, called spacetime Q-Network (STQN), that exploits predictive spatiotemporal encoding to reliably learn highly uncertain environment. The proposed method consists of two primary components. The first component is the successor representation with theta phase precession implements hippocampal spacetime encoding, acting as a rollout prediction. The second component, called Q switch ensemble, implements prefrontal population coding for reliable reward prediction. We also implement a single learning rule to accommodate both hippocampal-prefrontal replay and synaptic homeostasis, which subserves confidence-based metacognitive learning. To demonstrate the capacity of our model, we design a task array simulating various levels of environmental uncertainty and complexity. Results show that our model significantly outperforms a few state-of-the-art RL models. In the subsequent ablation study, we showed unique contributions of each component to resolving task uncertainty and complexity. Our study has two important implications. First, it provides the theoretical groundwork for closely linking unique characteristics of the distinct brain regions in the context of RL. Second, our implementation is performed in a simple matrix form that accommodates expansion into biologically-plausible, highly-scalable, and generalizable neural architectures.

 bioRxiv Subject Collection: Neuroscience

 Read More

Leave a Reply

%d bloggers like this: