Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Auditory category learning emerges from multiple neurocomputational systems
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
Qiren Dong1, Chen Hong2, Gangyi Feng3; 1Chinese University of Hong Kong
Auditory category learning requires the brain to transform continuously varying acoustic input into discrete, behaviourally meaningful categories. This process provides a controlled test case for speech category learning, in which similar sound-to-category mappings must be acquired under greater acoustic and linguistic complexity. A central unresolved question is whether auditory category learning is supported by a single dominant computational mechanism or by multiple interacting neurocomputational systems that contribute distinct representational operations throughout learning. Here, we addressed this question by combining source-resolved MEG with multilayer representational similarity analysis (RSA) to track how learning-dependent changes in deep neural network representations align with cortical learning dynamics during auditory category learning. Participants learned two novel ripple auditory categories while MEG activity was recorded. We compared neural representational geometry with representational changes across feedforward (convolutional neural network, CNN), recurrent, and contextual (transformer) neural network architectures. This comparison served as a mechanistic test of different computational operations. Feedforward models probed hierarchical acoustic-feature transformation, recurrent models probed temporally accumulated, category-relevant updating, and contextual models probed distributed, global representational organization across cortical systems. By relating model changes across training checkpoints to human neural representations across space and time, we tested both shared and architecture-specific computational contributions to newly formed auditory category representations. Learning-dependent brain–model alignment increased progressively across post-stimulus windows, with the strongest source-level effects emerging in the 400–600 ms window. Overall source-level RSA revealed significant and spatially distributed cortical alignment across architectures (mean network-level Fisher-z RSA approaching r ≈ 0.02, FDR-corrected q < .05), indicating that multiple neural-network architectures captured distributed aspects of cortical representational geometry during auditory category learning. Source-level cortical maps further revealed architecture-specific spatial organization: CNN and Transformer models showed broad distributed cortical alignment patterns, with Transformer models exhibiting the strongest overall alignment across learning stages, suggesting strong correspondence with globally shared cortical representational geometry. By comparison, recurrent architectures showed more selective and learning-sensitive cortical organization. In the model-only partial RSA analyses, all four architectures exhibited significant positive unique variance contributions across source-level network representations (BH-FDR corrected within net17 space), indicating partially dissociable computational contributions beyond variance shared across models. However, Long Short-Term Memory (LSTM) models consistently explained the largest source-level unique variance across all post-stimulus windows, with the strongest effects emerging in the 400–600 ms window (peak unique R² ≈ 2.3 × 10⁻⁴), whereas CNN, GRU, and Transformer models showed smaller but reliably positive unique contributions. This dissociation between overall RSA and partial RSA suggests that Transformer representations primarily captured broadly shared cortical representational structure, whereas recurrent architectures contributed more architecture-specific temporal updating dynamics during auditory category learning. Together, these findings support a multi-system account of auditory category learning. Sound-to-category mapping does not arise from a single learning mechanism, but from distinct computational systems that contribute different representational operations across time and cortex. More broadly, this study shows that controlled auditory category learning, combined with time-resolved brain-model alignment, could reveal the computational architecture through which the human brain acquires categorical sound representations, providing a mechanistic bridge to speech category learning.
Topic Areas: Speech Perception, Computational Approaches