Compact and Low-Power Neural Spike Compression Using Undercomplete Autoencoders.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Engineering

Background:

• Implantable microsystems are crucial for collecting and transmitting neural data in neuroscience research.
• High data rates from neural recordings necessitate on-chip data compression to prevent damaging power dissipation.
• Existing methods may struggle to balance compression ratios with signal fidelity.

Purpose of the Study:

• To present a novel data compression system for brain-computer interfaces (BCIs).
• To achieve high compression ratios for neural spike data while maintaining signal quality.
• To design and implement the system using a standard CMOS process.

Main Methods:

• Utilized undercomplete autoencoders for neural data compression.
• Developed integrated spike detection and autoencoder-based compression modules.
• Implemented the system in a 45-nm CMOS process for post-synthesis simulations.

Main Results:

• Achieved an average spike reconstruction quality of 14-dB SNDR at 32× CR, 18-dB at 16× CR, 22-dB at 8× CR, and 35-dB at 4× CR.
• The compression module demonstrated low power consumption (1.4–222.5 μW/channel) and silicon area (0.018–0.082 mm²/channel).
• Performance varied based on the desired compression ratio and number of channels.

Conclusions:

• The proposed autoencoder-based system offers an effective solution for neural data compression in BCIs.
• The system achieves significant compression ratios with high signal-to-noise-and-distortion ratios (SNDR).
• The low power and area requirements make it suitable for implantable microsystems.