Decoding EEG signals to understand decision-making and learning

Abstract

Perceptual learning refines how sensory evidence is encoded and transformed into choices. This thesis examined how short-term training (three consecutive days) reorganizes behavior, EEG dynamics, and latent decision processes across visual, auditory, and audiovisual modalities. Participants performed a speeded face–car categorization task with trial-by-trial accuracy feedback. To track learning with millisecond precision, EEG was analyzed using single-trial linear discriminant analysis (LDA) in a response-locked, sliding-window framework. Time-resolved classification (Az) distinguished Day 1 from Day 3 activity and yielded discriminant amplitudes (Y-values) at pre-response (decision) and post-response (feedback) peaks. A neurally informed hierarchical drift diffusion model (HDDM) then linked trial-wise Y-values to drift rate (v), boundary separation (a), and non-decision time (t). Behaviorally, training improved accuracy and shortened reaction times (RTs), with steeper gains for visual and AV trials. Neurally, Az exceeded chance throughout the decision-to-feedback interval, with robust pre-response (~−200 ms) and post-response (~+100 ms) peaks that were largest for visual and AV conditions. Feedback-locked Y-values predicted faster RTs on the next trial across all modalities, with the strongest coupling in AV trials; decision-locked Y-values also scaled with performance. HDDM revealed a coherent learning signature: higher v, lower a, and shorter t on Day 3. Trial-wise Y-values positively modulated v and were associated with reduced a and t, indicating that feedback-related signals calibrate both evidence quality and decision policy. Together, these results show that short-term perceptual learning is supported by a temporally distributed, modality-general mechanism that enhances evidence accumulation, reduces decision caution, and streamlines sensory–motor processing, with multisensory contexts conferring the greatest efficiency benefits.