An Autonomous Brain Machine Interface

Abstract

The aim of our work has been to advance the field of brain machine interfaces by creating a novel algorithm that can be implemented in the real-world. Most systems currently use supervised learning to train their algorithms for a pre-programmed environment. In an endeavor to move beyond supervised learning strategies, our group has created the first reinforcement learning brain machine interface in macaques, where reward expectation is the driving force behind the value of an action. After achieving this milestone, we discovered that the primary motor and somatosensory cortices were modulated by reward and successfully trained a classifier to detect rewarding trials from non-rewarding trials. We then implemented this classifier in an actor-critic brain machine interface, where the primary motor cortex signals were used to critique the actions of the agent. This is the first non-human primate BMI where a scalar value derived from the activity of the primary motor cortex is used to evaluate the actions of the system.The long-term ramification is an effective human neuroprosthesis for everyday life.