A 128-Channel Extreme Learning Machine-Based Neural Decoder for Brain Machine Interfaces.

Area of Science:

• Neuroscience
• Computer Engineering
• Biomedical Engineering

Background:

• Current brain-machine interface (BMI) algorithms require significant power, limiting their practical application.
• Existing motor intention decoding methods are often PC-dependent, hindering portability and real-time use.

Purpose of the Study:

• To develop a low-power, efficient machine learning coprocessor for motor intention decoding in BMIs.
• To enhance the accuracy and robustness of neural decoding for BMI applications.

Main Methods:

• Implementation of a 0.35-μm CMOS machine learning coprocessor utilizing the Extreme Learning Machine (ELM) algorithm.
• Integration of low-power analog processing with a second-stage learning mechanism for enhanced robustness.
• Verification using neural data from monkey finger movement experiments and asynchronous neural spikes.

Main Results:

• Achieved an energy efficiency of 3.45 pJ/MAC at a 50 Hz classification rate.
• Demonstrated a 99.3% decoding accuracy for movement type in monkey experiments.
• Increased classification accuracy by 5% for decoding movement timing using time-delayed feature enhancement and reduced programmable weights by approximately 2X via sparsity promotion.

Conclusions:

• The developed coprocessor offers a power-efficient solution for real-time motor intention decoding in BMIs.
• The chip architecture enhances decoding accuracy and robustness, paving the way for more advanced BMI systems.
• Sparsity promotion and feature enhancement techniques contribute to improved performance and reduced hardware complexity.