Identifiable VAEs for Neural Latent Discovery
Author(s)
Wu, Zijing
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Advances in neural recording enable us to record hundreds of neurons simultaneously, placing an increasing demand of developing appropriate computational statistical tools to analyze high-dimensional neural data. In neural latent analysis, where the latent representation of the raw neural data is extracted to understand the internal state of the biological neural network, variational auto-encoder (VAE) and its variations are emerging as a set of promising approaches. Recently, pi-VAE (Zhou, 2020) was proposed to address the identifiability issue of applying Latent Variable Models to neural spike data. However, pi-VAE’s dependency on external label information and failure to meet identifiability requirements due to its discrete observation model have limited its applicability. In this paper, we present a novel method called dynamical identifiable VAE (di-VAE) to address these limitations. Our approach proposes a continuous relaxation of the multivariate Poisson distribution, eliminating the need for external labels and introducing temporal dynamics through the use of history latent information as the conditional variable u. We systematically explore different di-VAE configurations using a rat hippocampus place cell dataset and further validate our chosen implementation on the Neural Latents Benchmark datasets. The results demonstrate competitive qualitative and quanttatative performance.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis