Generative Neural Spike Prediction from Upstream Neural Activity via Behavioral Reinforcement

It is quite challenging to predict dynamic stimulation patterns on downstream cortical regions from upstream neural activities. Spike prediction models used in traditional methods are trained by downstream neural activity as the reference signal in a supervised manner. However, downstream activity is unavailable when neurological disorders exist. This study proposes a reinforcement learning-based point process framework to generatively predict spike trains through behavior-level rewards, solving the difficulty. The framework is evaluated to reconstruct the transregional spike communication during motor control through behavioral reinforcement. We show that our methods can generate spike trains beyond the collected neural recordings and achieve better behavioral performance.