Vector Quantized Representations for Efficient Hierarchical Delineation of Behavioral Repertoires

Tianqing Li, Ugne Klibaite, Jumana Akoad, Joshua Huang Wu, Timothy W DUNN

Primary Area: applications to neuroscience & cognitive science

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: animal behavior, neuroscience, unsupervised unit discovery

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

Abstract: Understanding animal behaviors and their neural underpinnings requires precise kinematic measurements plus analytical methods to parse these continuous, multidimensional measurements into interpretable, organizational descriptions. Existing approaches can identify stereotyped behavioral motifs, given 2D or 3D keypoint-based data but are limited in their interpretability, computational efficiency, and/or ability to seamlessly integrate new behavioral measurements. In this paper, we propose an end-to-end behavioral analysis approach that dissects continuous body movements into sequences of discrete latent variables using vector quantization (VQ). The discrete latent space naturally defines an interpretable deep behavioral repertoire composed of hierarchically organized behavioral motifs. Using recordings of freely moving rodents, we demonstrate that the proposed framework faithfully supports standard behavioral analysis tasks and enables a series of new applications stemming from the discrete information bottleneck, including realistic synthesis of animal body movements and cross-species behavioral mapping.

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: 8649