Memory guidance of value-based decision making at an abstract level of representation

Value-based decision-making often involves situations where the optimal choice isn't immediately clear. In such cases, past experiences stored in memory can significantly influence decisions. Memory guidance can be based on similarity (physically similar stimuli) or association (events linked through prior experience). Sensory preconditioning, where an association between two stimuli is learned before one stimulus is paired with a reward, demonstrates how value can spread through association. This study focuses on understanding the specificity of associative memory guidance. The representation format (abstract or specific) plays a critical role in determining the specificity of memory guidance. Abstract representations, encompassing information across multiple dimensions, are efficient and prevent memory overload, but can be prone to errors and overgeneralization. Conversely, specific representations, such as pictorial details, are flexible but demand high storage capacity and can be too narrow to guide decisions effectively in novel contexts. The research aims to determine whether memory guidance for faces uses abstract or pictorial representations by examining the transfer of value within a face category using a sensory preconditioning paradigm. The study will address two primary questions: 1) Does memory guidance operate at the level of individual stimuli or entire categories? 2) Is memory guidance pictorial (view-specific) or abstract (view-invariant)? The study also investigates the roles of exposure duration, awareness of associations, individual memory capacity, and image-computable memorability.

Existing research highlights the role of memory in guiding value-based decisions. Studies indicate that the hippocampus plays a crucial role in bridging past experience with future decisions. Sensory preconditioning, a classic example of associative learning, demonstrates how value can spontaneously transfer between associated stimuli, even without direct reward pairing. However, the level of specificity at which this memory guidance operates remains unclear. Some studies suggest category-level generalization, while others emphasize the importance of specific, pictorial details. Different memory systems utilize both abstract and specific representations, coexisting with varying temporal decay functions. The visual system, for instance, shows both view-dependent and view-invariant representations. The question of whether memory guidance for faces relies on pictorial or abstract representations is addressed by this study, capitalizing on the well-defined and high-dimensional nature of faces, which have both specific and abstract representations in the brain.

Thirty healthy volunteers (mean age 26.9 ± 5.7 years, 17 female) participated in a three-phase sensory preconditioning experiment. Phase 1 (Association Learning): Participants were exposed to 12 grayscale images depicting distinct male Caucasian faces, grouped into six pairs (S1, S2). One group received one block (12 trials per pair) and the other group received four blocks (48 trials per pair) of training. A cover task ensured attention to stimuli without focusing on associations. Phase 2 (Reward Learning): Participants underwent classical conditioning, where half of the S2 stimuli were paired with monetary reward (S2+), and the other half with a neutral outcome (S2−). Phase 3 (Decision-Making): Participants performed a 2-alternative forced choice (2AFC) task. They chose between S2+ vs. S2− and S1+ (previously paired with S2+) vs. S1− (previously paired with S2−). To test for generalization, new views of S1+ and S1− (S1*) were also presented. Decision bias, normalized by S2 bias, measured the transfer of reward. Reaction times were also recorded. Additional measures included: likeability ratings (before and after the association phase), an explicit memory test for associations, and the Cambridge Face Memory Test (CFMT) to assess face episodic memory capacity. Image-computable memorability scores were obtained using MemNet, a deep neural network. Statistical analyses included t-tests, ANOVAs, and correlations to examine decision bias, generalization, and the influence of various factors.

The results demonstrated successful conditioning, with participants significantly preferring S2+ over S2−. Crucially, participants also showed significant transfer of reward from S2 to the original views of S1 (S1+), replicating the sensory preconditioning effect. Importantly, this reward transfer generalized to new views of S1 (S1*), indicating that memory guidance was not image-specific but operated at the level of facial identity. There was no significant difference in decision bias between original and novel views of S1, suggesting view invariance. Analysis of reaction times indicated that processing resources for S1 and S2 were similar. Furthermore, decision bias was not significantly affected by likeability or image-computable memorability. The amount of training during the association phase also did not significantly affect memory-guided decision-making or generalization. Explicit memory tests revealed that participants were not consciously aware of the S1-S2 associations, suggesting automatic memory guidance. Face episodic memory capacity (assessed by CFMT) did not correlate with decision bias, indicating distinct capacity constraints for memory guidance compared to episodic memory.

This study demonstrates that associative memory guidance of value-based decision-making is both specific enough to operate within a category (faces) and flexible enough to generalize across different views of the same identity. The findings challenge the notion that such memory guidance is strictly pictorial or solely category-level. Instead, it suggests an operation at a meaningful level of abstraction that neither over- nor underspecifies associations, balancing efficiency and accuracy. The generalization observed across head orientations extends previous research on stimulus generalization using unidimensional stimuli. This extends generalization to associated, non-rewarded stimuli, and demonstrates that it can occur along abstract dimensions. Brain regions involved in generalization—including category-selective visual cortex, hippocampus, orbitofrontal cortex, and striatum—may play a role in this process. While the hippocampus may represent higher-order relationships, view-invariant representations in higher visual cortex and orbitofrontal cortex could support generalization across head orientations. The absence of a correlation between generalization and image memorability further supports the conclusion that decision-making was driven by reward association rather than inherent stimulus properties.

This research reveals a flexible and adaptive memory system guiding value-based decisions. Memory guidance effectively balances specificity with generalization, operating at a level of abstraction that optimizes decision-making efficiency. Future studies should investigate this flexibility with stimuli lacking ecological relevance and explore the precise neural mechanisms responsible for this flexible memory system. Further research should also consider expanding the stimuli used to test the breadth of this generalization ability.

While the study controlled for several factors, it is possible that other, uncontrolled, stimulus dimensions might have influenced the results. Furthermore, the statistical power to detect small effects related to the amount of training was limited by the sample size. Lastly, this study focused on faces, which have inherent ecological relevance. Extending these findings to other stimulus categories will increase the generalizability of the findings.