Poster Presentation

Poster B42 in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Dynamic noise correlations between feature-selective brain regions support efficient learning

Joonhwa Kim¹, Caesar Dai¹, Apoorva Bhandari¹, Matthew Nassar¹; ¹Brown University

Presenter: Joonhwa Kim

Learning to assign credit to relevant features is a challenging problem. Recent theoretical work has demonstrated that stimulus-independent covariation in trial-to-trial neural activity fluctuations may reflect a dimensionality reduction mechanism that facilitates faster and more robust learning, but this proposal has yet to be empirically verified. Here we performed an empirical test of this theory by combining computational modeling, functional magnetic resonance imaging, and behavioral measures. Forty-five human subjects performed a task requiring discrimination of multi-dimensional perceptual stimuli while undergoing fMRI. In each block of trials, subjects learned a classification boundary in a two-dimensional feature space (color and motion) that was translated occasionally and unpredictably, forcing them to learn continuously on one of two possible joint feature dimensions (color+motion or color-motion, depending on block). A two-layer feedforward neural network performing this task showed that amplifying gain along task-relevant dimensions produced noise correlations aligned to these dimensions and enhanced learning. Human behavioral data was best fit by models with higher gain on task-relevant dimensions, suggesting that subjects focused on and updated the joint-feature dimensions that were most important for learning and task performance, and raising the question as to whether humans would show dynamic modulations of noise correlations across blocks to shift learning onto the relevant task dimension akin to gain adjustments in the model. Concurrently collected fMRI data revealed that noise correlations across feature-sensitive regions are modulated in a task-dependent manner, in line with this basic model prediction. Specifically, when residual fMRI data were passed through decoders trained on color and motion respectively, noise correlations between the resulting feature noise vectors were more positive during [color+motion] learning blocks and more negative during [color-motion] learning blocks, supporting the idea that the brain dynamically amplifies task-relevant dimensions to improve learning.

Topic Area: Methods, Tools, Theory & Neural Coding

Extended Abstract: Full Text PDF