Morning session

Memory search has typically been studied using small word lists (e.g. in free recall) or limited categories of words (e.g. in semantic fluency tasks). However, memory search in real-world contexts is much less constrained, potentially spanning thousands or millions of distinct items. This poses a challenge for modeling, predicting, and understanding memory processes in more naturalistic settings. We propose a novel approach that combines large language model (LLM) embedding methods for extracting features of arbitrary items with linear models of memory dynamics for searching over vast spaces of items based on these features. The resulting models are highly flexible and scalable, and can be tractably fit on large amounts of data. The parameters recovered from these fits can be used to interpret processes, and predict both established and new patterns involving contextual influences and complex dynamics. We demonstrate the value of our approach by discussing its applications to various domains in cognitive psychology, including free association, active learning, analogical reasoning, counterfactual reasoning, decision making, and open-ended thought. Our work highlights the potential of combining LLM embeddings with computationally tractable memory models to advance our understanding of unconstrained memory search in real-world contexts.

This is an in-person presentation on November 21, 2024 (09:10 ~ 09:40 EST).

It has long been hypothesized that episodic memory supports adaptive decision making by enabling mental simulation of future events. Yet, attempts to characterize this process are surprisingly rare. On one hand, memory research is often carried out in settings that are far removed from ecological contexts of decision making. On the other hand, models of adaptive choice only invoke episodic memory in highly stylized terms, if at all. In this talk, I will present TCM-SR, a novel process-level model that grounds model-based evaluation in empirically informed dynamics of episodic recall. In this model, the probability of retrieving each available memory is governed by the Successor Representation (SR), a biologically plausible world model in reinforcement learning. The evolution of these probabilities based on past retrievals, in turn, is dictated by the Temporal Context Model (TCM), a prominent model of episodic retrieval. Through a series of model simulations, I will show that the patterns of episodic retrieval suggested by this model enables flexible computation of decision variables. On this basis, a number of previously described features of episodic memory might serve an adaptive purpose in sequential decision making. For instance, the contiguity effect enables mental simulation via model-based rollouts, and emotional modulation improves generalization and decision efficiency when limited experience is available. By bridging computational models across memory and decision-making domains, we generate theoretical and empirical predictions linking episodic memory to adaptive choice in sequential tasks.

This is an in-person presentation on November 21, 2024 (09:40 ~ 10:10 EST).

Our experience of the world occurs in 3D, yet it is unknown how the rich sense of depth contributes to memories for events. In this talk, I will present a novel paradigm using immersive virtual reality (VR) to investigate 3D memories. Participants experienced real-world video events from a first-person perspective while wearing an immersive VR headset. The sense of depth was manipulated by presenting videos to each eye offset by the horizontal separation between the eyes (i.e., binocular disparity) and compared to a condition in which identical videos were presented to each eye, which I will refer to as 3D and 2D, respectively. Across two behavioural studies, we examined how 3D experiences influenced the subjective experience of remembering. In Study 1, we found that 3D memories were associated with higher vividness and scene-related aspects of remembering. In Study 2, we replicated these findings when memory was tested following a 24-hour delay. We adapted our paradigm to investigate the neural mechanisms that support the formation of 3D memories, using an MRI compatible VR headset. Our initial findings indicate that the 3D memories were associated with greater subsequent memory effects in the hippocampus. Together, this research suggests that dimensionality influences how people remember real-world events and demonstrates the potential of immersive VR for investigating the neural mechanisms that support the formation of real-world memories.

This is an in-person presentation on November 21, 2024 (10:30 ~ 11:00 EST).

What makes memories for realistic events more robust than memories for unstructured stimuli like word lists? In two current studies, we are investigating how durable event memories are scaffolded by schematic maps, including familiar spatial contexts. We hypothesize that there are two critical factors in this scaffolding process: the quality of the map itself and the creation of a strong conjunctive association that binds an event to the map. In our first study, participants learned to navigate a 23-room "museum" in immersive virtual reality and were later asked to remember an item placed into each room. We found that stability of spatial maps *before* being exposed to the items predicted subsequent success in neural reinstatement of the items; specifically, rooms with more reliable representations served as better anchors for remembering items. In our second study, participants were trained for four weeks in the "memory palace" memorization technique, in which they imagined a sequence of vivid events to associate items with locations along a spatial route. As they built their expertise in this technique, we found that that their events grew more detailed and creative, with both their description of the events and their neural representations of the events becoming more distinct from the isolated items and locations. These findings suggest that memories are strongest when they are include a large number of context-specific details, and that context is part of well-learned cognitive map.

This is an in-person presentation on November 21, 2024 (11:00 ~ 11:30 EST).