Visual Memory: Encoding, retrieval, capacity

Talk Session: Monday, May 18, 2026, 10:45 am – 12:15 pm, Talk Room 2
Moderator: Ming Meng, UAB

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

Talk 1, 10:45 am, 42.21

Revisiting Iconic Memory with a New Experimental Paradigm

Gal Vishne^1,*, Miranda Ye^2,*, Nicholas B. Turk-Browne³, Michael N. Shadlen^1,#, John Morrison^1,2,#; ¹Columbia University, ²Barnard College, Columbia University, ³Yale University

We experience the visual world as rich and detailed, yet working-memory studies suggest we retain only a few items at a time. Sperling's seminal 1960s experiments provided the first evidence for a higher capacity short-term buffer, later dubbed "iconic memory." In those studies, participants were flashed a 4x3 letter array and were cued after its disappearance to report only one row, which substantially improved performance relative to a full-report condition. Despite the theoretical importance of this topic, research on iconic memory has slowed, partly due to limitations of the original paradigm: reporting multiple items leads to output interference; identifying letters amplifies processing interference; some letters are easier to identify than others; and there is no way to measure the precision of memories. Here, we introduce a new paradigm designed to overcome these limitations. On each trial, participants were flashed (50 ms) nine colored dots evenly spaced on a circle (4° radius). For uniform perceptual spacing, colors were sampled from an equiluminant circle in perceptual CIELAB space (center a=20, b=38, radius=60). After a variable delay (0–1000 ms), the fixation changed color to match one of the dots, and participants made an eye-movement to the location of that color. This design offers several advantages, among them: one response per trial avoids output interference; as low-level continuous features, colors create less processing interference and support estimates of memory precision. We found that across trials overall performance declined with array-to-probe delay. Modeling error patterns as a mixture of uniform and von-Mises distributions (following Zhang & Luck, 2008) revealed two distinct mechanisms behind this decay: increase in guessing rate and a reduction in memory precision. These findings demonstrate the prevalence of very short-term, high-capacity visual memory across tasks and provide a new platform for a revival of research on iconic memory.

Talk 2, 11:00 am, 42.22

Semantic understanding, not just coherent object structure, strengthens visual working memory

Yong Hoon Chung¹ (), Sam Jung¹, Timothy Brady², Viola Störmer¹; ¹Dartmouth College, ²University of California San Diego

What constrains visual working memory capacity? Recent work shows that memory for real-world objects is markedly better than for abstract stimuli (e.g., colored squares or scrambled objects; Chung et al., 2024). This raises an important, yet unanswered, question: is this advantage driven by perceptual differences between these stimuli – real-world objects have cohesive form structure and are perceptually more distinct than abstract stimuli – or does it instead reflect the incorporation of semantic knowledge into visual working memory? Here, we test this by comparing memory performance and neural maintenance activity for real objects with counterfeit objects generated by generative adversarial networks (Cooper et al., 2023). Importantly, we confirmed with norming studies that counterfeit objects preserved the object form and visual similarity structure of their real counterparts, only varying in their recognizability. If the real-object advantage is purely perceptual, then matching objects on form and similarity should preserve memory benefits. In contrast, if semantic knowledge contributes to storage, then disrupting recognizability while preserving visual structure should reduce memory performance and reduce neural maintenance signatures. Strikingly, enhanced memory performance emerged only for real objects, with counterfeit objects resulting in the same performance level as scrambled control stimuli (Exp. 1, N = 50). Human electrophysiological results further revealed increased neural engagement for real-world objects, expressed in greater CDA amplitudes and more stable patterns over time across encoding and maintenance periods (Exp. 2, N = 24). Finally, correlation analyses (Exp. 3, N = 300) showed that subjective familiarity of each object predicted memory for real objects, whereas low-level features, specifically colourfulness, predicted memory for counterfeit objects, suggesting distinct mechanisms underlie memory performance across stimulus types. Overall, these data show that visual working memory – long viewed as a primarily sensory system – is strongly shaped by higher-level semantic structure for complex and naturalistic stimuli.

Talk 3, 11:15 am, 42.23

Visual Pattern Completion of Occluded Objects Promotes Robust Visual Learning

Rosemary Cowell¹, Jennifer Gove¹, David Huber¹; ¹University of Colorado Boulder

Recognizing an occluded object is effortless: visual cortex fills in the missing parts to infer a recognizable whole. Across six behavioral experiments and an fMRI study, we tested whether part-to-whole visual pattern completion constitutes not just a retrieval operation, but a learning mechanism that strengthens object representations. The behavioral experiments employed a “peephole” learning paradigm: first, participants studied whole objects, then each object was assigned either to retrieval practice, cued by a circular aperture offering a peephole view, or to restudy. At final test, retrieval-practiced objects were remembered better than restudied objects – a visual testing effect. The task minimized the utility of verbal/semantic information, suggesting that this testing effect occurs without semantic mediation, contrary to several influential theories. Furthermore, when retrieval no longer required pattern completion—because the peephole cue was removed—the testing effect disappeared, suggesting that pattern completion was responsible for learning. This provides the first evidence for a non-semantically mediated testing effect for visual stimuli, driven by occlusion-induced pattern completion. To test directly whether pattern completion in visual cortex underpins this novel testing effect, we took the task into the fMRI scanner. We recorded brain activation during peephole-cued recall. We also ran “template scans,” in which participants viewed whole objects. We trained a classifier to map template patterns from visual cortex onto object labels (e.g., lawnmower, clock, car). Next, we took “recall” brain patterns elicited by peephole cues (e.g., of a clock) and asked the classifier to predict the item being recalled. Item prediction was above chance, indicating that visual cortex performs pattern completion during peephole-cued recall. Moreover, classifier accuracy was higher for retrieval-practiced items than restudied or novel items. We conclude that visual pattern completion underpins the observed testing effect and serves a broader function than recognizing occluded objects, providing a potent mechanism for visual learning.

Talk 4, 11:30 am, 42.24

Pushing the envelope: Active viewing drives boundary extension

Hong B Nguyen¹ (), Benjamin van Buren¹; ¹The New School

When we actively move to view a scene, planning those movements requires predicting what we will see next. Sensorimotor predictions lie at the core of the “active learning” advantage, wherein observers better remember information if they can control what they see, and when they see it. Compared to passive viewing, active viewing may increase memory precision because it better aligns incoming visual input with the observer’s current preparatory and attentional state. Here, we asked whether visual predictions generated during active viewing not only enhance memory for what was seen, but also bias memory toward details which were never visible. In Experiment 1, observers viewed scene images by either using their mouse to actively move a small viewing aperture, or passively watching a replay of another participant’s viewing trajectory. After a 3-s viewing period, they completed a recognition memory test in which they discriminated the final view visible through the aperture from a decoy view which was slightly zoomed-in or zoomed-out. Active viewers showed robust boundary extension, more frequently selecting zoomed-out decoys than zoomed-in decoys; passive viewers showed the opposite pattern, consistent with boundary contraction. If active viewing drives greater boundary extension due to motor predictions, active viewers should specifically misremember having seen information in the direction their viewing trajectory was headed. In Experiment 2, we displaced the decoy view forward or backward relative to the participant’s last heading direction. Active viewers’ memory was biased toward unseen information in front of their viewing trajectory, and this bias was stronger than for passive viewers. Thus active exploration drives memory toward predicted future views—not just any unseen content. Together, these experiments demonstrate that sensorimotor predictions cause us to actively fill in what lies beyond a visible boundary. We conclude that active viewing constructs richer, yet more inferential, memory representations than passive viewing.

Talk 5, 11:45 am, 42.25

Robustness of prediction based false memory

Mahveen Salman Mubarak¹ (), Saba Halabisaz¹, Olya Bulatova², Caroline Yuan¹, Keisuke Fukuda^1,2; ¹University of Toronto Mississauga, ²University of Toronto

To make sense of the world in real time, the human brain predicts incoming visual events. These predictions, based on learned environmental regularities, enable efficient integration of visual information despite inevitable neural processing delays. Prior work has shown that events that confirm or violate our predictions are well remembered (Atienza et al., 2011; Bein et al., 2021). However, it remains unclear how predictions that are never confirmed or violated influence visual memory. To test this, we had participants perform a color memory task in which they first saw a gray silhouette of a real-world object and were asked to explicitly predict what color it would become using a continuous color wheel. After the prediction, the silhouette either turned into 1) the same color as the prediction (confirmation), 2) the opposite color (180° offset, large violation) or 3) a similar color (45° offset, small violation) to their prediction thus rebutting their prediction, or 4) remained colorless, thus neither confirming or rebutting their predictions. Critically, participants were instructed to remember the actual color of the object, not their prediction. Subsequent memory recall revealed that making an explicit prediction was sufficient to elicit a false memory of the predicted outcome when their prediction was not confirmed. Furthermore, direct rebuttal of the prediction did not eliminate this prediction-based false memory, even with a large violation. A series of follow-up experiments revealed that the prediction-based false memory cannot be explained by action selection (Experiment 2) or pre-existing schemas (Experiment 3), and it is robust against overnight decay (Experiment 4), unlike the memory of the actual outcome. Together, these findings demonstrate the strong influence of explicit predictions on what we remember, underlining the powerful role of expectations in shaping episodic memory.

This research was supported by the Natural Sciences and Engineering Research Council (RGPIN-2024-05727).

Talk 6, 12:00 pm, 42.26

The factors that determine trial-by-trial memory behavior across 6 million trials

Wilma Bainbridge¹ (); ¹University of Chicago

It is well known that many factors influence memory performance, including participant factors (e.g., individual differences, memory capacity), stimulus factors (e.g., memorability, distinctiveness), and contextual factors (e.g., local image context, trial order). However, little work has compared how important each factor is in determining memory. To test this question, I analyzed 6,223,953 memory decisions made by 13,946 participants on a continuous recognition memory task with 26,107 object images. I examined how 12 different factors commonly discussed in the memory field (e.g., memorability, image context and distinctiveness, previous trial performance, repetition lag, trial number, etc) are predictive of trialwise memory performance. These factors combined explained over half the variance in memory behavior (61.2%). The participant and the image’s memorability were the most impactful determinants of memory, with little variance captured by contextual factors, like the experimental distinctiveness of an image. The impact of memorability remained strong even when modeling memorability as predictions from a neural network (ResMem) independent from the experimental context, suggesting this is truly a stimulus-specific effect, rather than something about the stimulus set. Next, I ran 1000 mini-experiments within the data, testing whether the influence of these factors changes when looking at more homogenous or heterogeneous participants or contexts. More diverse participants resulted in less of an impact of context but an unchanged impact of memorability. More diverse images also did not change the impact of memorability on memory performance. Finally, I tested the factors that contribute to false alarms, and discovered that these are influenced by a wider variety of factors (intrinsic false alarmability, but also contextual and participant factors). In sum, this work demonstrates that surprisingly a lot of memory is predetermined and context-independent. I advocate for more models integrating stimulus, participant, and contextual factors in conjunction.

W.A.B. was supported by the Alfred P. Sloan Foundation and the National Science Foundation under CAREER Grant No. 2441710.