TI-VAE: A temporally independent VAE with applications to latent factor learning in neuroimaging

Eloy Geenjaar, Amrit Kashyap, Noah Lewis, Robyn Miller, Vince Calhoun

Keywords: variational autoencoder, computational neuroscience, latent factor analysis, latent factor learning, fMRI, sequential variational autoencoder, somatomotor cortex, weight sharing, inductive bias

Abstract: Functional magnetic resonance imaging (fMRI) data contain complex spatiotemporal dynamics, thus researchers have developed approaches that reduce the dimensionality of the signal while extracting relevant and interpretable dynamics. Recently, the feasibility of latent factor analysis, which can identify the lower-dimensional trajectory of neuronal population activations, has been demonstrated on both spiking and calcium imaging data. In this work, we propose a new framework inspired by latent factor analysis and apply it to functional MRI data from the human somatomotor cortex. Models of fMRI data that can perform whole-brain discovery of dynamical latent factors are understudied. The benefits of approaches such as linear independent component analysis models have been widely appreciated, however, nonlinear extensions are rare and present challenges in terms of identification. Deep learning methods are potentially well-suited, but without adequate inductive biases with respect to spatial weight-sharing may heavily overparameterize the model for the dataset size. Due to the underspecification of neuroimaging approaches, this increases the chances of overfitting and picking up on spurious correlations. Our approach extends temporal ICA to the non-linear case and generalizes weight sharing to non-Euclidean neuroimaging data. We evaluate our model on data with multiple motor sub-tasks to assess whether the model captures disentangled latent factors corresponding to each sub-task. Then, to evaluate the latent factors we find further, we compare the spatial location of each latent factor to the known motor homunculus. Finally, we show that our latent factors correlate better to the task than the current gold standard of source signal separation for neuroimaging data, independent component analysis (ICA).

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)

TL;DR: Our approach extends temporal ICA to the non-linear case and generalizes weight sharing to non-Euclidean neuroimaging data.