| Invited Symposia | Symposia

Note: Symposia have not yet been scheduled.

Representational change in the hippocampus: How task structure, development, and learning objectives impact differentiation and integration

Chairs: Samuel Lippl¹, Anna Schapiro²; ¹Columbia University, ²University of Pennsylvania
Presenters: Michael Mack, Serra Favila, Alison Preston, Samuel Lippl

To adapt rapidly to new environments, we need to both distinguish between related events and generalize across them. The hippocampus plays an important role in both of these abilities, by differentiating and integrating related inputs. What determines whether hippocampal representations differentiate or integrate? How can the hippocampus do both at once? In this symposium, we bring together experimental and theoretical perspectives that address these questions. Dr. Michael Mack will present a category learning task that requires both learning a general structure and memorizing exceptions to that structure and will show that different pathways within the hippocampus support integration and differentiation. Dr. Serra Favila will discuss how hippocampal representations may be shaped not only by our current task structure but also by our expectation of how useful integrated or differentiated memories will be for future tasks. Dr. Alison Preston will show how development shapes the interplay between integration and differentiation, and how the ability to coordinate between these two operations develops. Finally, Samuel Lippl will present a computational model that explains how different learning objectives shape integration and differentiation under associative learning rules, providing a novel theoretical account for repulsion, an extreme form of differentiation.

Presentations

When to integrate, when to differentiate: Category exception learning as a probe of hippocampal function

Michael Mack¹; ¹University of Toronto

Understanding the hippocampus's role in learning requires paradigms that can push the system toward both integration and differentiation, since real-world learning demands both: extracting what experiences share and encoding how new instances may depart from it. Over the past decade, we’ve explored how category exception learning offers exactly this leverage. As learners build concepts from experiences that share regular features, the hippocampus must integrate across instances; when exceptions arrive that violate these regularities, reconciling them instead requires differentiating the exception from the regularities it defies. I will present two fMRI studies that use this paradigm to trace how integration and differentiation demands are met across levels of hippocampal organization. First, by combining diffusion-weighted imaging with fMRI to isolate the functional "footprints" of the monosynaptic (MSP) and trisynaptic (TSP) hippocampal pathways, we show that MSP-footprint activation tracks regularity learning early in training, while TSP-footprint activation emerges with exception encoding later, linking each pathway to one side of the integration-differentiation functional toolkit. Second, we examined hippocampal subfield representations directly, asking whether resolving an exception drives differentiation from, or integration with, competing category representations. We found that learning pushed exception and regular-item representations toward differentiation in dentate gyrus and integration in CA1, especially when exceptions were confusable with a competing category. Notably, the degree of this reorganization was associated with individual learners' success. Together, these studies show that category exception learning is a productive empirical lens on how the hippocampus flexibly shifts between integrative and differentiating computations.

Memories reflect the expected value of abstraction

Serra Favila¹; ¹Brown University

Memories of past experience are critical for making adaptive decisions. However, different decision-making contexts require different amounts of memory abstraction. Specific memories are valuable when it is essential to get the details exactly right, whereas abstract memories are valuable when it is useful to generalize knowledge across situations. While there is evidence that both memory formats exist in the mind and brain, it is unclear what mechanism controls which format is expressed. We propose that memory abstraction is determined by contextual beliefs about how useful abstraction will be for future decisions, i.e. the expected value of abstraction (EVA). Across two behavioral experiments, we tested whether EVA can account for memory format and its downstream utility for decision-making. In Experiment 1, we manipulated EVA by providing participants with explicit priors about whether abstraction would be useful before they learned a network of associations. In Experiment 2, we manipulated EVA through trial-and-error learning about the importance of indirect associations in the network. Across both experiments, a higher EVA led to more false alarms on a memory test. However, when participants were presented with decisions requiring them to mentally navigate the network, a higher EVA made them more likely to succeed at problems querying long-range indirect associations. Both of these effects were well-explained by a computational model of memory abstraction in which a higher EVA leads to longer-range integration. Planned work will test whether EVA can explain context-dependent differences in whether the hippocampus expresses differentiated or integrated representations.

Development expands the scale and flexibility of hippocampal representational learning

Alison Preston¹; ¹University of Texas at Austin

Adaptive memory requires two seemingly opposing operations: differentiating similar experiences to preserve event-specific detail and integrating across experiences to extract shared structure. How the developing hippocampal system learns to balance these demands remains unclear. I will present neuroimaging evidence that development transforms representational learning across associations, events sequences, and spatial environments. When experiences overlap, children form differentiated hippocampal representations that support high-fidelity retrieval of individual events but leave relationships across events to be reconstructed through effortful retrieval. With development, representational learning becomes capable of preserving event-specific information while also integrating commonalities, allowing adults to access both individual memories and their relational structure. This expansion is also evident in the domain of time. Children form local, forward-directed links between temporally adjacent events, whereas maturation broadens the temporal scale of hippocampal integration and supports flexible, bidirectional representations of temporal structure. Spatial learning reveals a related transformation in how hippocampal representations adapt to change. Children differentiate new events from those previously experienced in the same location, remapping spatial information into differentiated, high-dimensional hippocampal representations. Adults instead compress experience around behaviorally relevant commonalities while preserving distinctions needed for precise spatial decisions. Together, these findings show that development does not simply strengthen memory or replace differentiation with integration. Rather, it expands the capacity to coordinate both operations, enabling event details and shared structure to be represented simultaneously and deployed flexibly. Across these domains, we further show how maturing interactions between hippocampus and frontoparietal cortices may expand representational capabilities and support flexible expression during memory-guided behavior.

An analytic theory of hippocampal representation learning

Samuel Lippl¹; ¹Columbia University

Flexible behavior requires both integration across related stimuli and differentiation between them, processes that both implicate the hippocampus. While there are various theories of how the hippocampus may shape representations to accomplish these goals, there is an enduring unsolved mystery in this field: hippocampal representations sometimes become more dissimilar for related cues than unrelated cues—a counterintuitive phenomenon known as "repulsion." Repulsion has been suggested to be inconsistent with classical associative learning rules, which are commonly understood to promote integration of related stimuli. Here we present a simple computational model of the hippocampus that accounts for repulsion using standard learning rules. The model consists of two learning streams corresponding to the monosynaptic and the trisynaptic pathway that are trained jointly. We develop an analytic theory of this model that lets us pinpoint the exact mechanism giving rise to repulsion, which is analogous to an adaptation aftereffect in vision. Crucially, our theory predicts a dissociation between the two pathways, with integration in the monosynaptic pathway and differentiation or repulsion in the trisynaptic pathway. It further predicts that repulsion should arise either for cues with a shared associate or for cues that are themselves more similar to each other, accounting for key observations in the literature. Our model thus replicates experimental findings of repulsion and establishes a novel, parsimonious account of representational shaping in the hippocampus.