tong-wu-umn/spike-compression-autoencoder

Deep Compressive Autoencoder for Large-Scale Spike Compression

Deep Compressive Autoencoder for Large-Scale Spike Compression

The repository contains source codes for the paper Deep Compressive Autoencoder for Action Potential Compression in Large-Scale Neural Recording accepted by Journal of Neural Engineering.

Pytorch 0.4.1 is required to run the model. An Nvidia GPU with over 4GB memory is preferred.

Authors

Tong Wu¹, Wenfeng Zhao¹, Edward Keefer², Zhi Yang^1*

¹ Department of Biomedical Engineering, University of Minnesota, Minneapolis, MN, 55455, USA

² Nerves Incorporated, Dallas, TX, 75214, USA

Datasets

Three datasets are used in the paper, which are all publicly available and can be downloaded at the following links:

• Wave_Clus
• HC1
• Neuropixels (requires permission from Dr. Nick Steinmetz if used for publication)

Spikes can be extracted from the datasets either according to their ground truth timing (if available) or using detection methods.

Alignment of spikes is required before compression.

Spikes need to be organized into dimensions (Num_batch x Num_channel x Num_pionts) before loading into CAE, where Num_channel is M_spk (typically equal to or smaller than the number of recording channels), Num_points is the length of each spike.

Data should be in .mat format and put in the folder ./data (not contained in the repository; needs to be created by users).

How to use

spk_vq_cae.ipynb is the top-level notebook.

Under the section Parameter Configuration:

• spk_ch is M_spk, by default 4.
• spk_dim is Num_points.
• vq_num is the number of vector quantization codes.
• cardinality is the number of groups in a CNN layer, by default 32.
• dropRate is the rate of dropout, by default 0.2.
• batch_size is the batch size, by default 48.

Under the section Preparing data loaders:

• For non-denoising autoencoder, noise_file and clean_file refer to the same dataset.
• For denoising autoencoder, noise_file refers to a distorted version of clean_file.