Schedule of Events | Search Abstracts | Symposia | Invited Symposia | Poster Sessions | Data Blitz Sessions

Symposia

Hippocampal predictions link perception and memory

Symposium Session 2: Sunday, April 14, 2024, 1:30 – 3:30 pm EDT, Ballroom Center

Chairs: Peter Kok¹, Morgan Barense²; ¹University College London, ²University of Toronto
Presenters: Dhairyya Singh, Anna Schapiro, Oliver Warrington, Mariam Aly, Oded Bein

Memory and perception are intimately linked through prediction. Memory is used to generate predictions of upcoming perceptual events, and mismatches between those predictions and perception determine how we update our memories. Despite this, memory and perception are traditionally studied separately, with knowledge gained in one field not significantly impacting the other. In this symposium, we will bring together researchers that bridge this gap by studying predictions in order to uncover the interplay between what we see and what we remember. We will discuss how perceptual predictions influence memory formation (Singh, Bein) and how memory-based predictions influence sensory processing (Kok, Aly). The presented research will span the gamut of cognitive neuroscience techniques, from psychophysics and computational modeling to cutting edge human neuroimaging. The speakers also span a range of career stages, from PhD and postdoc to PI. Following the individual talks, we will have a general discussion on the current state of knowledge of the field and the most pressing outstanding questions (chaired by Barense). With this symposium, we aim to use prediction as a means to bridge the gap between memory and perception, allowing researchers in the two fields to interact and learn from each other, promoting new understanding and novel research avenues.

Presentations

Statistical learning drives predictive shifts in object memory

Dhairyya Singh¹, Anna Schapiro¹; ¹University of Pennsylvania

The world around us is highly regular, with objects clustering predictably in time and space. How do statistical relationships between the objects in our environment impact the way we remember them? Previous work has shown that neural representations of objects morph closer together if they co-occur reliably in time, and there is a forward bias in these shifts, with objects moving closer to their successors in the hippocampus. It remains unclear whether and how these representational shifts manifest in our memory and what learning mechanisms may underlie them. Employing color memory as a continuous index of memory change, we investigated how the strength and direction of transition statistics shape object memory over the course of statistical learning. Participants viewed sequences of colored shapes consisting of pairs, with the first member of the pair predicting the second at varying transition probability strengths. The shapes from each pair were assigned nearby colors in a perceptually uniform colorspace. We found that color memory for predictive shapes systematically distorted towards the color of their successor shapes over the course of learning, especially for pairs with highly reliable transition probabilities. Our model of the hippocampus, trained through simple auto-encoding on an analogous task, recapitulated human behavioral results, with predictive object representations asymmetrically distorting toward their successors and the magnitude of the shift increasing in proportion to transition probability. Together, our results show how our memory for objects systematically shifts toward the predictable future and how simple hippocampal learning mechanisms can give rise to these changes.

Learning and communication of perceptual predictions by the hippocampus

Oliver Warrington¹; ¹University College London

We constantly exploit the statistical regularities in our environment to help guide our perception. The hippocampus (HC) has been suggested to play a pivotal role in both learning environmental statistics, as well as exploiting them to generate perceptual predictions. However, the mechanisms whereby the hippocampus learns such predictions remain unclear, as does its potential role in communicating predictions to sensory cortex. Here, we present the results of high-resolution human fMRI work directly investigating this. We collected submillimetre 7T fMRI data to measure layer-specific activity in the medial temporal lobe (MTL), specifically the entorhinal (EC), parahippocampal (PHC), and perirhinal cortex (PRC). Layer-specific fMRI allows one to infer the direction of communication between the HC and MTL, since superficial layers of MTL project to the HC, while deep MTL layers receive feedback projections from HC. Participants performed a task in which an auditory cue predicted shapes in 75% of trials. Crucially, we omitted the expected shape on 25% of trials, thus isolating the prediction signal from the bottom-up input. Activity patterns in the posterior subiculum reflected the predicted-but-omitted shapes. We used layer-specific informational connectivity analysis to ask: In which direction are predictions communicated between the HC and neocortex? We find that the CA2/3 subfield of the hippocampus sends information specific to the predicted-but-omitted shapes to the deep layers of the PHC. These findings suggest that the HC sends prediction signals to the neocortex, adding weight to the suggestion of its pivotal role in generating perceptual predictions.

Hippocampal memories enable preparation for anticipated attentional goals

Mariam Aly¹; ¹Columbia University

In the complex world around us, it can be difficult to know what to pay attention to. We often solve that problem by leveraging our memories of past experiences, which help us determine which part of the world is most relevant for our goals. One key advantage of using memory to guide attention is that it affords the opportunity to predict what we will likely have to pay attention to in the future. What mechanisms allow memories to enable preparatory coding for anticipated attentional goals? A key candidate region for coordinating preparation for memory-guided attention is the hippocampus, given its roles in attention, long-term memory, and prediction. In complementary fMRI studies, we investigated how the human hippocampus may enable preparation for upcoming attentional states. First, we show that the hippocampus exhibits preparatory coding for anticipated attentional tasks. This hippocampal preparation signal is stronger when attention is guided by memory rather than an explicit instruction. Second, we show that the ability of the hippocampus to differentiate similar memories allows the resolution of competition in memory-guided attention. Hippocampal differentiation of competing memories predicts both the precision of subsequent memory-guided eye movements and the precision of preparatory coding in visual cortex prior to the execution of visual search. Together, these studies show that hippocampal memories can serve an adaptive role in online attentional behavior by influencing our ability to predict how and where we should direct our attention.

Building event models: how hippocampal subfields differentially segment events through learning, and how anxiety alters event segmentation

Oded Bein¹; ¹Princeton University

Event representations are knowledge of unfolding occurrences within a context. They guide context-relevant predictions, memory, and behavior, and are segmented at changes of context, termed ‘event boundaries’. I ask how the hippocampus represents events through learning, and how event representations alter in anxiety. Neurotypical participants viewed sequential events while undergoing high-resolution fMRI. Event lists were repeated in identical order to induce learning and predictability. Behavioral segmentation, indexed by increased reaction times at event boundaries, was observed in all repetitions, which is at odds with prediction errors at boundaries triggering segmentation. Neurally, as events became learned, hippocampal CA3 multivoxel activation patterns clustered to reflect the event context. In contrast, the dentate gyrus exhibited event-specific temporal pattern separation, with temporally proximal items belonging to the same event becoming more differentiated, potentially maintaining event details. While these results might reflect appropriate context representation, variations in context representation may play a role in mental health concerns. In anxiety, overgeneralization of fear might result from poor separation between contexts. However, recent evidence showed enhanced sensitivity to context changes in anxiety. In large online studies, we asked how trait anxiety influences event segmentation. We found that, during story reading, anxious individuals (based on self-report) read slower specifically sentences that included boundaries. In another study, we tested how participants segment movie clips into events. Anxious individuals segmented events more like the group norm, potentially reflecting more precise segmentation. Together, these studies suggest hierarchical event representations in the hippocampus, and slower, but more precise, event segmentation in anxiety.