Embedding Learning for Approximating Person-specific Cognitive Similarity

Yujin Cha

Keywords: Psychological embedding, Metric learning, Similarity, Cognitive representation, Autoencoder, Medical image

TL;DR: Modeling Psychological Embeddings to Improve Person-Specific Similarity Inference with Autoencoders

Abstract: Metric learning is often applied in scenarios where labels are well-defined or where there is a ground truth for semantic similarity between data points. However, in expert domains such as medical data, where experts perceive features and similarities differently on an individual basis, modeling psychological embeddings at the individual level can be beneficial. Such embeddings can predict factors that influence behavior, such as individual uncertainty, and support personalized learning strategies. Despite this potential, the amount of person-specific behavioral data that can be collected through similarity behavior sampling is insufficient in most scenarios, making modeling individual cognitive embeddings challenging and underexplored. In this study, we proposed integrating supervised learning on small-scale similarity sampling data with unsupervised autoencoder-based manifold learning to approximate person-specific psychological embeddings with significantly improved similarity inference performance. We conducted a large-scale experiment with 121 clinical physicians, measured their cognitive similarities using medical image data, and implemented person-specific models. Our results demonstrate that even in complex expert domains, such as medical imaging, where cognitive similarity varies between individuals, person-specific psychological embeddings can be effectively approximated using limited behavioral data.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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: 6599