Most of what we know about neural mechanisms of incremental learning through feedback comes from descriptive, univariate analyses. Here, we go one step further, seeking brain activity that is not just statistically reliable (potentially small but significant) but can track such learning at the item level, taking a classifier-based approach to narrow in on basic neural encoding processes. Participants ( \(N=45\) ) learned 48 word-value mappings through trial-and-error. First, we checked whether established EEG markers of feedback processing, the feedback-related negativity (FRN) and frontal midline theta activity (FMT), are in fact predictive of trial-to-trial learning of the current item—and they were (above chance, but not by much), validating the behavioural relevance of those features. Next, we asked whether there might be considerably more information about encoding on single trials beyond these statistically robust, regular signals. Indeed, multivariate classifiers (LDA and SVM), incorporating signal-features beyond the FRN and FMT, predicted learning more substantially and exceeded previous performance on episodic recognition using the same basic approach (Chakravarty et al., Journal of Neurophysiology, 124(6), 2060–2075, 2020). Time-frequency spectral features produced better classifications (AUC \(\sim \) 0.7) than time-domain features. Finally, a possible shortcut due to accuracy varying systematically with trial number could not explain away classification success. In sum, FRN and FMT are not just descriptive of feedback-driven learning but also a bit predictive—but are the tip of the iceberg (subject-specific, spatiotemporal features) uncovered by the multivariate classifiers. This extends current classifier-based approaches to brain activity from episodic memory to incremental, feedback-driven learning.
