Understanding Scalp Electrophysiology via Machine Learning: Insights into Working Memory and Credit Assignment
Date:
This PhD exit seminar presented research exploring how machine learning techniques can enhance our understanding of scalp-recorded electrophysiological signals in the context of working memory and decision-making.
The talk began by reviewing canonical event-related potential (ERP) components, traditionally studied using univariate methods, and introduced multivariate decoding approaches to improve detection and interpretation of these signals. In the second section, I examined whether the P3b component reflects enhanced working memory representations, drawing on converging evidence from behavioral data, univariate ERP analysis, and neural decoding.
The final section shifted focus to contingency learning, applying machine learning to both scalp voltages and alpha-band power in a two-armed bandit paradigm. I tested whether working memory representations were reactivated at feedback to support credit assignment for unsustained items, revealing evidence of memory reinstatement and decision-linked reactivation.
This work demonstrates the value of machine learning in electrophysiological research and contributes to ongoing efforts to decode the contents and timing of working memory from noninvasive neural data.