| Invited Symposia | Symposia

Note: Symposia have not yet been scheduled.

Invited Symposium 4 - From Prediction to Explanation: Opportunities and Challenges for NeuroAI

Chair: Leila Wehbe¹; ¹Stanford University
Presenters: Leila Wehbe, Mariya Toneva, Anna Ivanova, Benjamin Cowley

Rapid advances in artificial intelligence have transformed our ability to predict neural activity and compare representations in artificial and biological systems. At the same time, neural and cognitive data offer an opportunity to modify artificial models by providing new training signals, architectural constraints, and benchmarks for human-like computation. Yet predictive accuracy or improved brain alignment alone does not guarantee that a model captures the computations underlying human cognition. Moving NeuroAI from prediction toward explanation will require models that can learn from rich neural data, generalize across people and experimental settings, distinguish among cognitive processes, and ultimately yield interpretable and testable accounts of neural computation. This symposium brings together researchers developing complementary approaches to this challenge, spanning naturalistic human neuroimaging, brain-guided training of artificial models, cognitive analyses of brain–AI alignment, and model-driven studies of neural populations. The talks will highlight opportunities to use neural measurements not only to evaluate AI models, but also to constrain and improve them, while addressing fundamental questions about what successful brain prediction actually tells us about cognition and mechanism. Together, the symposium will emphasize the need to move beyond benchmark prediction toward models that are generalizable, cognitively meaningful, and experimentally interpretable.

Presentations

RABBiT: Rapidly adaptive BOLD foundation model via brain-tuning

Mariya Toneva¹; ¹Max Planck Institute for Software Systems

Language understanding in the brain is context-dependent, varying across experimental stimuli and individuals, which makes it difficult to build computational models that generalize across both. This calls for a foundation model of language-evoked brain activity that can capture shared structure while adapting efficiently to new participants and inputs. We introduce RABBiT (Rapidly Adaptive BOLD foundation model via BraIn-Tuning), a compact audio-to-fMRI encoder designed for accurate zero- and few-shot prediction. A comprehensive evaluation on 324 participants across multiple unseen fMRI datasets shows that RABBiT enables accurate zero-shot prediction of fMRI responses to natural speech across auditory and language-selective regions, surpassing the SOTA foundation model for fMRI and predictions based on group averages. With as little as 10 minutes of participant-specific data, RABBiT further improves performance via parameter-efficient tuning, substantially outperforming per-participant linear models. RABBiT's performance is driven by two key innovations: (1) learned region-specific attention, and (2) a decomposition of brain responses into shared and subject-specific components, combined with a brain-tuned speech backbone. In addition to supporting strong predictive accuracy, the structured, region-specific representations that RABBiT learns enable interpretability. By eliminating the need for extensive per-participant data and model fitting, RABBiT enables scalable population-level analyses of language in the human brain.

Predictive modeling of language-evoked brain responses

Anna Ivanova¹; ¹Georgia Tech

Stories evoke a rich array of activity throughout the brain. These widespread responses reflect the power of language to convey many different meanings, encourage reflection and reminiscence, and evoke imagery of specific scenes, events, and people’s mental states. Here, I propose a framework for dissociating specific types of language-evoked activations (e.g., physical scene simulation), language-specific activations (extracting meaning from words and phrases), and sensory activations (e.g., auditory cortex responses to spoken stories). I then show that localizing functional brain areas in individual participants allows us to dissociate representations that underlie brain responses in specific brain areas, identify confounds, and build more precise, interpretable models of story understanding in the brain.

Learning High-Resolution Brain Activity from Naturalistic MEG and fMRI

Leila Wehbe¹; ¹Stanford University

Current non-invasive neuroimaging techniques trade off spatial and temporal resolution: magnetoencephalography (MEG) captures rapid neural dynamics, while functional magnetic resonance imaging (fMRI) provides better spatial localization. Integrating these signals into a unified, high-resolution representation remains challenging, particularly for single-trial naturalistic data. We collected whole-head MEG as participants listened to more than seven hours of narrative stories also used in an open fMRI dataset (LeBel et al., 2023). We developed a transformer-based encoding model that jointly predicts MEG and fMRI across subjects while learning a latent representation of cortical activity with high spatiotemporal resolution. The model predicts MEG better than standard single-modality encoding models and, in simulations, recovers sources with higher spatial and temporal fidelity than minimum-norm solutions. Critically, the learned source representation generalizes across unseen subjects and recording modalities: it predicts electrocorticography (ECoG) responses in an independent dataset better than an ECoG-trained encoding model. By integrating large-scale naturalistic experiments, complementary neuroimaging modalities, and neural encoding models, this approach provides a practical route toward millisecond-and-millimeter brain mapping.

Compact deep neural network models of the visual cortex

Benjamin Cowley¹; ¹Cold Spring Harbor Laboratory

A powerful approach to understand the computations carried out by the visual cortex is to build models that predict neural responses to any arbitrary image. Deep neural networks (DNNs) have emerged as the leading predictive models, yet their underlying computations remain buried beneath millions of parameters. Here we challenge the need for models at this scale by seeking predictive and parsimonious DNN models of the primate visual cortex. We first built a highly predictive DNN model of neural responses in macaque visual area V4 by alternating data collection and model training in adaptive closed-loop experiments. We then compressed this large, black-box DNN model, which comprised 60 million parameters, to identify compact models with 5,000 times fewer parameters yet comparable accuracy. This dramatic compression enabled us to investigate the inner workings of the compact models. We discovered a salient computational motif: compact models share similar filters in early processing, but individual models then specialize their feature selectivity by ‘consolidating’ this shared high-dimensional representation in distinct ways. We examined this consolidation step in a dot-detecting model neuron, revealing a computational mechanism that leads to a testable circuit hypothesis for dot-selective V4 neurons. Beyond V4, we found strong model compression for macaque visual areas V1 and IT (inferior temporal cortex), revealing a general computational principle of the visual cortex. Overall, our work challenges the notion that large DNNs are necessary to predict individual neurons and establishes a modeling framework that balances prediction and parsimony.