Search Abstracts | Symposia | Slide Sessions | Poster Sessions
A unified neuro-computational framework for modeling category learning dynamics
Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai
Hanlin Wu1, Zhenjiang Cui1, Shuguang Yang3,4, Suiping Wang3,4, Gangyi Feng1,2; 1Department of Linguistics and Modern Languages, The Chinese University of Hong Kong, Hong Kong SAR, 2Brain and Mind Institute, The Chinese University of Hong Kong, Hong Kong SAR, 3School of Psychology, South China Normal University, Guangzhou, China, 4Philosophy and Social Science Laboratory of Reading and Development in Children and Adolescents (South China Normal University), Ministry of Education, Guangzhou, China
<INTRODUCTION> Human learning unfolds over time. Learners often encounter stimuli, make decisions, receive feedback, update internal representations, and gradually improve. Yet most computational accounts remain tied to specific domains or mechanisms. This limitation is evident in auditory category learning, a foundational process for speech acquisition. Existing theories emphasize cortico-striatal reinforcement learning, associative/statistical learning, Bayesian adaptation, or multiple learning systems. Each explains important aspects of learning, but no single architecture captures group-level trajectories, individual variability, item difficulty, and neural representational dynamics within one framework. Here, we introduce a compact transformer-based neuro-computational model that treats learning as a sequence of behavioral events and tests whether its internal states align with human brain representations. Although evaluated in the context of auditory category learning, this framework provides a general approach to modeling how learners change through sequential experience. <METHODS> We studied behavior and fMRI data from 134 adults learning two ripple-sound categories across six feedback-based blocks. A 2-layer, 4-head, 128-dimensional GPT-2 model was trained with next-token cross-entropy loss. A customized tokenizer encoded each learner’s trajectory as a sequence, enabling the model to predict categorization and response speed from prior trials using 10-fold cross-validation. Evaluation involved next-trial prediction (NT) using the full history, and autoregressive simulation (AR) where only part of the history was provided, with the model generating the rest. To compare the model and brain, per-trial hidden states from each layer were used in representational similarity analysis (RSA) with item- and subject-level models. <RESULTS> Under NT, the model closely captured group-level learning trajectories (ACC/RT: r = 0.99/0.98), subject variabilities in performance (r = 0.95/0.87), learning outcome (r = 0.87/0.84), and item variability in difficulty (r = 0.83/0.35; all p < .001). Importantly, the model also forecasted future learning from sparse individual history. With 3 blocks or even 1 block of learning history, AR successfully predicted categorization accuracy at the trajectory level (AR3/AR1: r = 0.99/1.00), subject level (r = 0.89/0.57), outcome level (r = 0.72/0.34), and item-difficulty level (r = 0.85/0.83; all p < .001). Model-brain alignment further revealed a layer-by-network gradient. The embedding layer was dominated by somatomotor (50%) and cingulo-opercular (22%) networks, whereas contextualized Layer 1/2 broadened to frontoparietal (Layer 1/2: 15%/17%), language (10%/12%), and default-mode (10%/12%) networks. Individual-differences RSA was significant for both layers, with clusters in bilateral anterior insula, pre-SMA, medial PFC, and left IFG operculum, almost entirely within frontoparietal and cingulo-opercular networks (Layer 1/2: 98%/99%). <CONCLUSION> These findings show that a compact transformer trained only on behavioral sequences can recover the multiscale dynamics of human learning and generate neurally meaningful representations of individual learners. By linking trial-level prediction, autoregressive learner simulation, item-difficulty modeling, and whole-brain representational alignment, this framework moves beyond specific models toward a scalable architecture for learning science. The same approach could be extended to speech, language, perceptual, motor, decision-making, educational, and clinical learning, as well as any domain where learning unfolds through sequential experience. More broadly, it offers a framework for forecasting individual learning outcomes, designing personalized training protocols, and testing mechanistic hypotheses in silico.
Topic Areas: Computational Approaches, Language Development/Acquisition