Deep compressive autoencoder for action potential compression in large-scale neural recording.

Area of Science:

• Computational Neuroscience
• Machine Learning for Neural Data
• Biomedical Engineering

Background:

• Large-scale neural recordings are crucial for understanding brain computations.
• High data transmission bandwidth and power consumption are major hurdles for neural interfaces.
• Reducing data rates before transmission is a key solution for scalable neural recording.

Purpose of the Study:

• To develop a deep learning-based compression model for multichannel action potentials.
• To reduce the data rate of neural signals for efficient transmission.
• To enable high-bandwidth, high-precision, large-scale neural interfaces.

Main Methods:

• Developed a deep compressive autoencoder (CAE) with discrete latent embeddings.
• Encoder uses residual transformations and vector quantization (VQ) for feature extraction.
• Decoder reconstructs spike waveforms from quantized latent embeddings via stacked deconvolution.

Main Results:

• The CAE model achieved compression ratios of 20-500×, outperforming conventional methods.
• Demonstrated high reconstruction accuracy and robustness against waveform variations and misalignment.
• Estimated hardware performance supports thousands of channels with low power consumption.

Conclusions:

• The proposed CAE model effectively reduces data transmission bandwidth for large-scale neural recordings.
• Maintains good signal quality, facilitating power-efficient and lightweight wireless neural interfaces.
• Open-sourced code implementation is available for further research and development.