Task-relevant autoencoding enhances machine learning for human neuroscience

Aug 3, 2026

Upload by Finn

Human neuroscience uses machine learning to uncover low-dimensional neural representations linked to behavior, yet current approaches often depend on large training sets and therefore overfit limited-sample neuroimaging data with high dimensionality. The TRACE framework leverages the principle that sought features are defined by behavioral relevance. TRACE is evaluated against standard and variational autoencoders and PCA on two truncated machine learning datasets, then benchmarked on fMRI from 59 subjects. TRACE improves classification accuracy and yields cleaner task-relevant representations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages41

Reading Time1 h 43 min

AccessFree to download

Introduction

Problem: limited samples and high-dimensional neuroimaging data

TRACE: Task-Relevant Autoencoder via Classifier Enhancement

Model assessment and benchmarking

Application to fMRI data

“Task-relevant autoencoding” enhances machine learning for human neuroscience Authors: Seyedmehdi Orouji 1 , Vincent Taschereau-Dumouchel2-3 , Aurelio Cortese4 , Brian Odegaard5 , Cody Cushing6 , Mouslim Cherkaoui6 , Mitsuo Kawato4 , Hakwan Lau7 , & Megan A. K. Peters 1,8 1 Department of Cognitive Sciences, University of California, Irvine, Irvine, California, USA 92697 2 Department of Psychiatry and Addictology, Université de Montréal, Montreal, Canada, H3C 3J7 . 3 Centre de recherche de l’institut universitaire en santé mentale de Montréal, Montréal, Canada. 4 ATR Computational Neuroscience Laboratories, Kyoto, Japan 619-0288 5 Department of Psychology, University of Florida, Gainesville, FL USA 32603 6 Department of Psychology, University of California Los Angeles, Los Angeles, 90095, USA 7 RIKEN Center for Brain Science, Tokyo, Japan 8 Center for the Neurobiology of Learning and Memory, University of California, Irvine, Irvine, California, USA 92697 Correspondence should be directed to: Seyedmehdi Orouji Department of Cognitive Sciences 2201 Social & Behavioral Sciences Gateway University of California, Irvine Irvine, CA 92697 [sorouji@uci.edu](sorouji@uci.edu) Megan A. K. Peters Department of Cognitive Sciences 2201 Social & Behavioral Sciences Gateway University of California, Irvine Irvine, CA 92697 [megan.peters@uci.edu](megan.peters@uci.edu) Abstract In human neuroscience, machine learning can help reveal lower-dimensional neural representations relevant to subjects’ behavior. However, state-of-the-art models typically require large datasets to train, so are prone to overfitting on human neuroimaging data that often possess few samples but many input dimensions. Here, we capitalized on the fact that the features we seek in human neuroscience are precisely those relevant to subjects’ behavior. We thus developed a Task-Relevant Autoencoder via Classifier Enhancement (TRACE), and tested its ability to extract behaviorally-relevant, separable representations compared to a standard autoencoder, a variational autoencoder, and principal component analysis for two severely truncated machine learning datasets. We then evaluated all models on fMRI data from 59 subjects who observed animals and objects. TRACE outperformed all models nearly unilaterally, showing up to 12% increased classification accuracy and up to 56% improvement in discovering “cleaner”, task-relevant representations. These results showcase TRACE’s potential for a wide variety of data related to human behavior. Keywords: human neuroscience, machine learning, dimensionality reduction, task-relevant representation, fMRI, MVPA, autoencoder 1. Introduction In studying the human brain and human behavior, we often use machine learning methods to home in on the (ideally lower-dimensional1–4) representations contained in multivariate, featurerich datasets. These data typically contain noisy, irrelevant signals 19–21 that we would like to filter out using methods such as multivariate decoders5–8 , various types of autoencoders, generative adversarial networks like InfoGAN9 , or even principal components analysis (PCA) 10–12. However, state-of-the-art machine learning methods typically require very large datasets to train while data for individual human subjects collected with methods such as functional magnetic resonance imaging (fMRI) 13–15 are often severely limited in sample size 16,17 (i.e. , have very few training exemplars compared to the dimension of data) . Consequently, these methods are susceptible to overfitting on such neuroimaging data, reducing their predictive power and utility 18–20. What’s more, parametric methods (such as PCA), which may better avoid the need for large training sets, by definition require rigid assumptions regarding the nature of the dimensionality reduction process and thus are limited a priori to insights consistent with these parametric assumptions. Thus, we are in need of a nonparamet

FAQs

Why do many machine learning models overfit human neuroimaging datasets?

They often require large training sets, but neuroimaging data for individual subjects are severely limited in sample size while having many input dimensions, increasing overfitting risk.

What is TRACE and how does it incorporate task information?

TRACE is a classifier-enhanced autoencoder that attaches a logistic regression classifier to the bottleneck layer, explicitly prioritizing features relevant to the subject’s behavioral task.

How was TRACE evaluated in the study?

TRACE was benchmarked using multiple quantitative metrics across different bottleneck dimensionalities on two truncated machine learning datasets, and then tested on fMRI data from 59 subjects.

Aug 3, 2026

Upload by Finn