Modelling mPFC Activities in Reinforcement Learning Framework for Brain-Machine Interfaces

Reinforcement learning (RL) algorithm interprets the movement intentions in Brain-machine interfaces (BMIs) with a reward signal. This reward can be an external reward (food or water) or an internal representation which links the correct movement with the external reward. Medial prefrontal cortex (mPFC) has been demonstrated to be closely related to the reward-guided learning. In this paper, we propose to model mPFC activities as an internal representation of the reward associated with different actions in a RL framework. Support vector machine (SVM) is adopted to analyze mPFC activities to distinguish the rewarded and unrewarded trials based on mPFC signals considering corresponding actions. Then the discrimination result will be utilized to train a RL decoder. Here we introduce the attention-gated reinforcement learning (AGREL) as the decoder to generate a mapping between motor cortex(M1) and action states. To evaluate our approach, we test on in vivo neural physiological data collected from rats when performing a two-lever discrimination task. The RL decoder using the internal action-reward evaluation achieves a prediction accuracy of 94.8%, which is very close to the one using the external reward. This indicates the potentials of modelling mPFC activities as an internal representation to associate the correct action with the reward.