Deep Independent Vector Analysis

Xinhui Li, Rogers F Silva, Vince Calhoun

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

Keywords: multimodal fusion, nonlinear IVA, MISA, iVAE

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a deep multivariate latent variable model, Deep Independent Vector Analysis (DeepIVA), for learning linked and identifiable disentangled representations across multiple data modalities.

Abstract: We introduce a deep multivariate latent variable model, Deep Independent Vector Analysis (DeepIVA), for learning linked and identifiable disentangled representations across multiple data modalities by unifying multidataset independent subspace analysis (MISA) and identifiable variational autoencoders (iVAE). DeepIVA aims to leverage hidden linkage information via the MISA loss to attain latent cross-modal alignment while leveraging the identifiability properties of the iVAE to ensure proper unimodal disentanglement. We propose a more strict set of performance measures, and demonstrate that DeepIVA can successfully recover nonlinearly mixed multimodal sources on multiple linked synthetic datasets compared with iVAE and MISA. We then apply DeepIVA on a large multimodal neuroimaging dataset, and show that DeepIVA can reveal linked nonlinear imaging sources associated with phenotype measures including age and sex.

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.

Submission Number: 8343