Probabilistic Tensor Decomposition of Neural Population Spiking Activity

Matlab (recommended) and Python (in developement) implementations of Soulat et al. (2021).

The model (A) decomposes an observed count tensor (eg. binned spikes) using a Negative Binomial distribution that depends on a shape parameter, a constrained offset (B) and low rank tensor (C). Variational inference is implemented using a Pólya-Gamma augmentation scheme.

Demo

To train the model(s) on the toydataset described in the paper open:

matlab/demo_vbgcp.m

Or:

python/examples/demo_tensor_variational_inference.ipynb

PG approximation Figures can be generated with:

matlab/study_polyagamma.m

Data Analysis

We process results from S.Keshavarzi (2021) https://doi.org/10.1101/2021.01.22.427789 and benchmark performance of our method compared to standard (G)CP baselines in terms of Variance Explained (A) Deviance Explained (B) and a robustness/similarity metric (C)

Figure generated using:

matlab/data_benchmark.m
matlab/data_benchmark_process.m

Citing us

If our work helps you in a way that you feel warrants reference, please cite the following paper:

@inproceedings{
soulat2021probabilistic,
title={Probabilistic Tensor Decomposition of Neural Population Spiking Activity},
author={Hugo Soulat and Sepiedeh Keshavarzi and Troy William Margrie and Maneesh Sahani},
booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
year={2021},
url={https://openreview.net/forum?id=1bBF5Zq1YHz}
}