Efficient in Vivo Neural Signal Compression Using an Autoencoder-Based Neural Network.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Engineering

Background:

• Conventional neural signal processing for brain-computer interfaces (BCIs) relies on spike counts, which is insufficient for continuous local field potential (LFP) data.
• High-density intracortical recordings in BCIs generate large neural data volumes, demanding high transmission rates.
• Efficient data transmission is critical for advancing LFP-based cognitive decoding in BCIs.

Purpose of the Study:

• To develop the first autoencoder-based digital circuit for efficient compression of in vivo LFP neural signals.
• To optimize the circuit for reduced computational complexity and memory requirements.
• To enable robust signal reconstruction for enhanced BCI performance.

Main Methods:

• Implementation of an autoencoder-based neural network for LFP signal compression.
• Algorithmic and architectural optimizations to minimize computational load and memory footprint.
• Design of an application-specific integrated circuit (ASIC) for the compression logic.

Main Results:

• The developed ASIC achieved the smallest silicon area and lowest power consumption among state-of-the-art compression ASICs.
• The circuit demonstrated a higher compression rate compared to existing methods.
• A superior signal-to-noise and distortion ratio was achieved, ensuring robust signal reconstruction.

Conclusions:

• The autoencoder-based compression circuit offers an efficient solution for transmitting LFP neural signals in high-density BCI applications.
• The designed ASIC represents a significant advancement in low-power, high-compression neural data processing.
• This technology has the potential to improve the feasibility and performance of LFP-based BCIs.