Visual Memory: Imagery, memorability, long-term

Talk Session: Tuesday, May 20, 2025, 8:15 – 10:00 am, Talk Room 2

Schedule of Events | Search Abstracts | Symposia | Talk Sessions | Poster Sessions

Talk 1, 8:15 am

What Makes Mental Images Vivid? Sharpness as the Key Visual Dimension

Xueyi Huang¹, Angela Shen², Emil Olsson², Kiarra Michelle I. Garcia², Nadine Dijkstra³, Megan A. K. Peters², Jorge Morales¹; ¹Northeastern University, ²University of California, Irvine, ³University College London

Visual mental images vary in their subjective vividness. Conventionally, this vividness has been measured on a unidimensional numerical scale. But how different visual qualities influence our experience of vividness is poorly understood. Here, we present results from a novel method for reconstructing visual properties of mental images as they vary across different subjective vividness ratings. On each trial, subjects saw a line drawing of an object and were instructed to imagine it in as much detail as possible. Then, they rated the vividness of their mental image on a 1-to-5 scale. Finally, a Voronoi tessellation pattern—a tiling of differently-colored shapes—appeared on the screen along with three sliders that controlled its sharpness, opacity, and saturation. Subjects abstracted the qualities of their mental image and applied them to the Voronoi pattern by adjusting the sliders. With a mixed-effects linear model that displayed high goodness of fit, we found that all three dimensions significantly predicted subjects’ vividness ratings. Importantly, at the group level, sharpness was the strongest predictor (followed by opacity and then saturation), and it also explained most of the random effects variance. At the subject level, sharpness was the top predictor of trial-by-trial vividness ratings as well. While our results showed individual differences in what visual features influence vividness ratings, the clear primacy of sharpness as the strongest and most common predictor can help guide further research into the visual properties and neural basis of mental imagery. Additionally, our modeling results were not correlated with the overall vividness of subjects’ mental imagery, suggesting that the pattern in which people relied on the three visual dimensions was independent of their overall mental imagery capacity. In conclusion, despite the subjective nature of mental imagery, these results show that our method can robustly reconstruct vividness ratings by quantifying the contributions of different visual features.

This work was supported by the Templeton World Charity Foundation, Inc. (funder DOI 501100011730) under the grant https://doi.org/10.54224/22032 to (JM, MAKP & ND) and in part by the Canadian Institute for Advanced Research (CIFAR, Fellowship in Brain, Mind, & Consciousness; to MAKP).

Talk 2, 8:30 am

The Art of Memory—Art contest reveals that artists can create memorable artworks, predicted by AI

Trent M Davis¹ (), Yifei Chen¹, Wilma A Bainbridge¹; ¹University of Chicago

Why do we remember some artworks but not others? Previous research has shown that paintings have an intrinsic memorability, where the same pieces are consistently remembered and forgotten across viewers (Davis & Bainbridge, 2023). Importantly, the memorability of famous paintings was significantly higher than non-famous paintings as predicted by ResMem, a neural network designed to predict the memorability of images without any outside information or context (Needell & Bainbridge, 2022). This was especially surprising given previous work showing that people are inconsistent in their ability to predict an image’s memorability (Revsine & Bainbridge, 2022). Inspired by these findings, we organized a national art contest with the following question in mind: Can artists intentionally create memorable and forgettable artwork? 87 artists from across the United States submitted to the contest, with about two thirds attempting to make memorable artwork and the remaining third, forgettable. Pieces were judged via a continuous recognition memory experiment (N=131), where we found entries to the memorable category were significantly more memorable than the forgettable category, suggesting that artists are able to manipulate the memorability of their artwork. We then created a hybrid art exhibit and in-person experiment where the 20 most memorable and forgettable artworks were displayed for a month at Connect Gallery in Chicago, IL. Viewers of the exhibit were invited to test their memory of the artworks (N=61), and we also tested people’s memories at a satellite event at VSS 2024 (N=140). Visitors in both samples remembered memorable pieces significantly better than forgettable ones (p<0.001), demonstrating artists were able to manipulate memory. Further, ResMem was significantly able to predict people’s memory performance both online and in-person (p<0.001). These results suggest that memory for artwork is manipulable and predictable, which has major implications for the design of both artwork and museums.

Talk 3, 8:45 am

Single neurons in human MTL track the depth-of-processing elicited by visual representations of images

Aalap Shah¹ (), Yuchang Tian¹, Qi Lin², Runnan Cao³, Shuo Wang³, Ilker Yildirim¹; ¹Yale University, ²RIKEN, ³Washington University in St. Louis

Decades of research have pointed to the human Medial Temporal Lobe (MTL) as the locus interfacing spontaneous visual processing and memory formation. The hippocampus, for instance, receives input from the ventral visual stream and encodes episodic memories. Yet, the algorithmic basis of how stimulus-driven visual processing modulates activity across MTL structures remains unclear. Here, we address this gap by leveraging computational techniques to model a unique dataset of single-cell recordings from the hippocampus (n=362 neurons) and amygdala (n=446 neurons) of humans (n=15) passively viewing real-world images. We hypothesize that a core psychological theory, the ‘depth-of-processing’ of visual inputs (Craik and Lockhart, 1972), explains neural processing across MTL structures. A recent study (Lin et al., 2024) introduced an image-computable signature of depth-of-processing: compression-based reconstruction error of visual representations obtained by training a linear sparse autoencoding model on the activations of a pre-trained convolutional neural network. However, the stimuli used in our neural dataset included several basic-level categories (e.g., animals, airplanes, buildings etc.) that proved challenging for the linear model to compress effectively. Therefore, we developed a higher-capacity, non-linear sparse autoencoding model trained on a large-scale, multi-category image dataset, ImageNet. We found that the same computational signature of Lin et al. (2024)—compression-based reconstruction error—positively correlated with the firing rates of hippocampus singe-cells under this more powerful model. No alternative model demonstrated this relationship with the hippocampus, including distinctiveness of visual representations and reconstruction error under a non-sparse variant of our model. Moreover, given the predominantly inhibitory connections from the hippocampus to amygdala; we predicted—and subsequently confirmed—a statistically significant, but notably negative correlation between reconstruction error and the firing rates of amygdala single-cells. Only our model captured this dissociative relationship across the two MTL structures: hippocampus and amygdala. These results suggest depth-of-processing as an algorithm-level account of stimulus-driven visual processing in the MTL.

Talk 4, 9:00 am

Memorability Beyond Semantic Features: Probing Memorability of Semantically Similar Images Through Generative Diffusion Model

Hyewon Willow Han^1,2, Yalda Mohsenzadeh^1,2; ¹Western Center for the Brain and Mind, Western University, London, Ontario, Canada, ²Vector Institute for AI, Toronto, Ontario, Canada

What makes an image more memorable? Recent findings suggest that semantic features exert a stronger influence than perceptual features on the memorability of object images (Kramer et al., (2023)). However, semantically similar images can still exhibit a diverse range of memorability. What features of an image make it more memorable among semantically identical images? To address this question, we propose a novel framework that generates cloned stimuli from original images, replicating both the image features and the memorability of the originals. We utilize the THINGS dataset (Hebart et al., (2023)), a naturalistic image collection comprising diverse kinds of object concepts and their associated memorability scores. First, we develop a memorability predictor model tailored to the THINGS dataset using CLIP (Radford et al., (2021)) and leverage recent advancements in large-scale generative diffusion models (Rombach et al., (2022)) to generate clones of these images. We demonstrate that the cloned images successfully preserved image attributes and features from the originals and exhibited comparable memorability scores. Subsequently, we further investigate what makes an image more memorable with the most forgettable and most memorable cloned images by analyzing groups of artificial neurons responsible for specific image features through the decomposition of image representations (Gandelsman et al., (2024)). Our findings reveal that cloned images not only successfully replicate the memorability of their originals but also exhibit variability in their memorability. By analyzing these variations, we gain deeper insights into perceptual features that drive image memory in addition to semantic features. This study highlights the potential of generative diffusion models to explore the cognitive and computational attributes underlying image memorability.

Talk 5, 9:15 am

Limited influence of sleep on drawings made from memory

Samuel R. Rosenthal¹, Emma Megla¹, Wilma A. Bainbridge¹; ¹Department of Psychology, University of Chicago

It is well known that memory worsens the longer it has been since encoding (Ebbinghaus 1885), and that memory is better after a period of sleep than the same period awake (Ellenbogen et al., 2006). While this is widely known to be true for recognition-based visual memory (Wagner et al., 2007), it is less clear for visual recall. To assess the effect of sleep on visual recall, we conducted two experiments. First, Prolific participants (N=188) encoded 4 scene images before undergoing a 10-hour delay, during which they were either awake or asleep. After the delay, the participants drew the 4 images from memory. Participants also encoded and immediately drew 4 scene images, to replicate prior effects without a delay (Megla et al., 2024). Following this, separate Prolific participants (N=393) rated which objects were present in the drawings. As expected, participants recalled significantly more images, and more objects from those images, immediately after encoding than after a delay. However, we surprisingly found no significant difference in memory quality between the sleep and wake conditions. In the second experiment, we increased interference between images by having participants memorize and draw scenes within the same category, which has been shown to result in worse memory quality (Hall et al., 2021). Participants (N=175) encoded 8 scene images, in which 4 were from the same scene category (“within”) while the remaining 4 were of different categories (“between”). Participants recalled significantly more images from the “within” scene category, but significantly more objects from the “between” scene category. However, despite successfully increasing interference between images, we found no significant difference between sleep and wake for either metric. These results challenge the commonly held belief that sleep benefits all types of memory and suggests that visual recall is not impacted by sleep.

Talk 6, 9:30 am

Contributions from Long-Term Memory Explain Superior Working Memory Performance for Meaningful Objects

Hyung-Bum Park¹ (), Edward Awh¹; ¹University of Chicago

Visual working memory (VWM) capacity has been claimed to be higher for meaningful objects compared to simple visual features, possibly due to their rich and distinctive representations. However, prior demonstrations have made this observation by comparing working memory performance for trial-unique objects and repeated sets of simple stimuli (e.g., a limited set of color categories). Unfortunately, this design includes a confound between meaningfulness and the strength of proactive interference, which is virtually absent for trial-unique object images. Thus, improved behavioral performance with meaningful stimuli could reflect contributions from episodic long-term memory (LTM) that are not accessible with repeated stimuli typically used in standard VWM capacity studies. To test this hypothesis, Experiment 1 measured VWM performance for trial-repeated colors, trial-repeated objects, and trial-unique objects. The results replicated the advantage for trial-unique objects over simple colors, but this advantage was eliminated with trial-repeated objects. Equivalent performance with colors and trial-repeated objects appears to contradict the claim that enhanced distinctiveness enables more meaningful objects to be stored in VWM. Instead, these findings indicate that LTM contributions in the trial-unique condition are eliminated by PI in the trial-repeated condition. To further test this interpretation, Experiment 2 measured contralateral delay activity (CDA), an electrophysiological marker of active VWM storage, for trial-repeated colors and trial-unique meaningful objects. Here again, we replicated the behavioral advantage for trial-unique objects, but CDA amplitudes plateaued at equivalent set sizes for objects and colors. Thus, our findings suggest that an equivalent number of trial-unique objects and colors can be stored in VWM, although testing with trial-unique objects invites contributions from familiarity-based representations in LTM (e.g., Endress & Potter, 2014; JEP:G).