The brain-machine interface (BMI) technology builds a communication pathway be-tween the brain and external devices, which provides a solution for disabled people to restore their lost function. Existing BMI frameworks generally let patients accomplish some pre-define tasks in the lab or hospital to build a decoding model. In a real application scenario, BMIs are expected to have the ability of autonomous learning because patients need to use the device to do many new tasks in their daily life. Since the brain has a developed learning mechanism, utilizing the learning related information from neural signals will advance the BMIs towards an autonomous learning system. Many studies have suggested that the medial prefrontal cortex (mPFC) is highly involved in reward-guided learning. Howeve...[ Read more ]

The brain-machine interface (BMI) technology builds a communication pathway be-tween the brain and external devices, which provides a solution for disabled people to restore their lost function. Existing BMI frameworks generally let patients accomplish some pre-define tasks in the lab or hospital to build a decoding model. In a real application scenario, BMIs are expected to have the ability of autonomous learning because patients need to use the device to do many new tasks in their daily life. Since the brain has a developed learning mechanism, utilizing the learning related information from neural signals will advance the BMIs towards an autonomous learning system. Many studies have suggested that the medial prefrontal cortex (mPFC) is highly involved in reward-guided learning. However, previous work mostly studied the functionalities of mPFC by averaging the signals across trials and analyzed on a univariate. The multivariate encoding analysis on a single trial basis has not been investigated. Since the implementation of BMIs requires high time resolution and real-time interaction between subjects and devices, it is also important to study how to utilize the encoded information to build a closed-loop autonomous learning BMI framework.

In this work, we conducted the neural dynamics analysis of mPFC activities and utilized the encoded information to improve the design of an autonomous BMI framework. Firstly, we did a multivariate encoding analysis on mPFC cortical signals to investigate what information mPFC activities encode during task learning. The result shows that mPFC activities simultaneously encode multiple information including goal planning, action execution and outcome evaluation. The encoded information is highly related to the audio cue in the task. Based on the encoding analysis, we designed a new behavior training system with more efficient audio feedback to facilitate subjects to learn the task. We observed a significant improvement of learning process and learning speed at the behavior level. Then we analyzed the neural response upon the audio feedback at both single neuron level and neuron population level. At last, we demonstrated the benefit of this analysis by using the encoding information to help the reward design of an autonomous BMI demo. Our results show that we can successfully extract the encoded information from mPFC activities and use the information to improve the design of autonomous BMIs. It reveals the potential of endowing BMIs with autonomous task learning ability utilizing subjects’ internal evaluation function.