A 192-Channel 1D CNN-Based Neural Feature Extractor in 65nm CMOS for Brain-Machine Interfaces.

Area of Science:

• Neuroscience
• Computer Engineering
• Biomedical Engineering

Background:

• Brain-Machine Interfaces (BMI) require efficient neural feature extraction for decoding user intent.
• Existing methods often face challenges with decoding stability, power consumption, and long-term performance.

Purpose of the Study:

• To present a novel 192-channel 1D convolutional neural network (1D CNN) based neural feature extractor for BMIs.
• To achieve state-of-the-art decoding stability with high power and area efficiency.
• To validate the performance and stability of the proposed feature extractor with human participants.

Main Methods:

• Developed a 192-channel 1D CNN architecture in 65nm CMOS technology.
• Implemented an on-chip model, FENet-66, supporting configurable layers and kernel lengths.
• Incorporated individually power-switchable channels and layers for optimized efficiency.
• Validated the system in real-time closed-loop experiments with human participants.

Main Results:

• Achieved state-of-the-art decoding stability at 1.8 μW and 12801 μm² per channel.
• FENet-66 demonstrated superior cross-validated decoding performance compared to existing feature sets.
• Features showed 18% higher average R2 decoding performance than Spiking Band Power (SBP), with 28% improvement in the 4th year.
• Custom 1D-CNN kernels outperformed wavelet features by 10% and compressed data by 38×.

Conclusions:

• The proposed 1D CNN architecture offers a scalable, power-efficient solution for neural feature extraction in BMIs.
• The FENet-66 model significantly enhances decoding performance and long-term stability for neural implants.
• This technology improves the viability of long-term neural implants compared to current low-power BMI hardware solutions.