Maximum correntropy based attention-gated reinforcement learning designed for brain machine interface

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Reinforcement learning (RL) is crucial for brain-machine interfaces (BMIs) to map neural activity to movement.
• Existing attention-gated RL (AGREL) improves efficiency but struggles with noisy neural signals.
• Outliers in neural data hinder accurate interpretation of neural-action relationships.

Purpose of the Study:

• To develop an enhanced AGREL algorithm robust to noise in neural signals.
• To improve the accuracy and efficiency of neural-action mapping in BMIs.
• To address the challenge of spatial credit assignment in complex BMI tasks.

Main Methods:

• An enhanced AGREL algorithm was developed using correntropy as a noise-insensitive criterion.
• The algorithm was tested on neural data from a monkey performing an obstacle avoidance task.
• Performance was evaluated by comparing success rates with the original AGREL.

Main Results:

• The enhanced AGREL algorithm demonstrated faster convergence during training.
• Success rates improved from 44.63% to 68.79% on average compared to the original AGREL.
• The correntropy criterion effectively reduced the impact of outliers in neural signals.

Conclusions:

• The combination of correntropy and AGREL offers a more robust approach for neural-action mapping in BMIs.
• This enhanced method improves performance by mitigating noise and outliers in neural data.
• The findings suggest a promising direction for developing more reliable and efficient brain-machine interfaces.