Special Symposia

Valence Influences on Episodic Memory: A Cognitive Processing Theory

The assumption that memory for emotional events (e.g., a graduation celebration or a funeral service) is superior to memory for routine, mundane events is widespread. However, evidence from episodic memory research provides, at best, mixed support for this assumption, leaving open questions about how and when emotions enhance memory. This line of research consisted of a meta-analysis along with three large-scale experiments, using memory tests that varied by the extent to which they relied on integrative/environmentally-driven versus elaborative/self-initiated processing. The main findings from this work validate the common belief that emotional events are remembered better than neutral ones, validating also more specific assumptions that the emotional enhancement effect is larger for negative than for positive pictures, larger in free recall than in recognition memory, and also larger after a long retention interval than a short one.

This work was guided by a novel theory I call Emotional Events Interrupt-Processing Augment (EEI/PA) theory. The EEI/PA theory predicted the main findings from my meta-analysis, as well as those of my experiments, all of which tested the EEI/PA’s core assumption: that valence influences encoding-phase accommodative/elaborative processing, but not assimilative/integrative processing. My work demonstrates that the EEI/PA theory can explain the findings based on the influence of valence and further highlights the need for an integrated account that identifies the unique contributions of both valence and arousal to the way in which we remember emotional events.

Modeling inter- and intra-individual variation in brain–valence relationships with deep generative models

Cognitive and affective neuroscience often relies on the assumption of a one-to-one, linear mapping between brain activity and behavior. This assumption, however, overlooks the principle of degeneracy, where multiple distinct neural states may produce the same psychological outcome. To address this, we developed a semi-supervised deep generative framework (Variational AutoEncoder with Classification head; VAE-C) that simultaneously models the distribution of brain states (unsupervised) and their relationship to behavioral labels (supervised). We then applied this framework to a densely sampled fMRI dataset (N=36) where participants viewed approximately 324 emotionally evocative videos over 4 separate sessions.

We found that brain–valence mappings are predominantly many-to-one, participant-dependent, and at times nonlinear. Model comparison revealed that for 31 of 36 participants, nonlinear classification models outperformed linear baselines, but that the strength of these effects varied considerably across participants. Furthermore, generative analysis revealed robust evidence of many-to-one mappings: hierarchical clustering across a range of dissimilarity thresholds showed that the overwhelming majority of participants had many-to-one brain-valence mappings (between 80% and 100% depending on threshold). When brain states were compared across individuals, many states were unique to a single individual or shared by only a small subset of individuals. Together, these findings underscore the limitations of seeking universal “neural signatures” for valence—and behavior more broadly—and suggest the necessity of modeling inter- and intra-individual differences in these processes.