Visual Memory: Neural mechanisms of working memory

Talk Session: Sunday, May 18, 2025, 10:45 am – 12:30 pm, Talk Room 1

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

Talk 1, 10:45 am

Inversion of feature preference in visual cortical neurons during working memory and attention

Diego Mendoza-Halliday¹ (), Andrii Zahorodnii², Haoran Xu², Christopher Cueva², Julio Martinez-Trujillo³, Robert Desimone²; ¹University of Pittsburgh, ²Massachusetts Institute of Technology, ³Western University, London, Ontario, Canada

Dominant theories and models of visual working memory are based on the principle that the maintenance of feature representations in working memory is subserved by the persistence of activity of neurons whose preferred feature closely matches the memorized feature. Similarly, dominant theories and models of attention are based on a feature-similarity gain principle—that the strength of attentional modulation of a given feature-selective neuron is determined by the similarity between its preferred feature and the attended feature. Here, we show experimental evidence that challenges these two principles. In macaque monkeys performing a working memory-guided feature attention task or delayed match-to-sample task for motion direction, we found that in nearly half of the neurons in visual cortical area medial superior temporal (MST), the preferred motion direction during the working memory delay period was opposite to the preferred direction during the stimulus presentation period. A similar inversion in feature preference was observed during feature attention in neuron from both MST and area middle temporal (MT), contrary to that predicted by the feature-similarity gain principle. Interestingly, along the visual processing stream, the percentage of neurons with inversions of feature preference during working memory or attention was highest in earlier processing stages, decreased downstream, and was lowest in later processing cortical areas, including the lateral intraparietal (LIP) and lateral prefrontal cortex (LPFC). These results challenge the generality of theories and models of working memory and attention that assume the predominant recruitment of neurons with preferred features that match the attended or memorized features. Last, dimensionality reduction analyses of neuronal population activity patterns during working memory in MST, and a comparison with competing computational models of working memory, suggest that the observed inversions of feature preference during working memory in visual cortical neurons represent a putative mechanism to protect working memory representations from sensory interference.

Talk 2, 11:00 am

Functional Relevance of The Medial Temporal Lobe in Visual Working Memory Quality Revealed by Within-Subject Lesion-Symptom Mapping

Weizhen Xie¹ (), Sanikaa Thakurdesai¹, Oceane Fruchet², Samantha Jackson², Radhika Chatterjee², Evalyn Johnson-Ramsay¹, Sara Inati², Zaghloul Kareem²; ¹University of Maryland, College Park, ²National Institute of Neurological Disorders and Stroke, National Institutes of Health

Classic lesion case-control studies suggest that the medial temporal lobe (MTL) has minimal involvement in visual working memory (VWM), particularly for simple features like colors and orientation gratings. However, recent findings from direct MTL recordings suggest its crucial role in VWM, potentially through pattern separation that reduces mnemonic interference during short retention intervals. The absence of significant findings in earlier studies may reflect less sensitive task measures or variability across individuals. To address these, we investigated the effects of MTL lesions on VWM representation using a color recall task in 40 neurological cases, assessing performance before and after neurosurgery through lesion-symptom mapping. Of these, 18 cases had MTL lesions, including the hippocampus, while 22 had no lesions or lesions outside the MTL, such as in the insula or prefrontal cortex. Measuring participants’ VWM recall variability and overall recall likelihood using a mixture model, we found that MTL lesions led to a significant increase in recall variability, indicating reduced VWM precision post-surgery within individuals. Finer lesion-symptom mapping revealed a strong correlation between hippocampal damage and recall variability, even after controlling for overall lesion size across the whole brain. However, while overall lesion size influenced recall likelihood, reflecting a smaller amount or quantity of remembered VWM content following brain resection in general, hippocampal lesion size could not account for this effect. These findings underscore the MTL’s specific role in supporting VWM quality, distinct from VWM quantity supported by distributed neocortical mechanisms. This dissociation challenges unitary models of VWM constraints that overlook the distinction between memory quality and quantity, emphasizing the need for updated frameworks incorporating critical neuropsychological evidence.

W.X. is supported by NIH grant R00NS126492

Talk 3, 11:15 am

Computational methods for extracting neural correlates of working memory and mental imagery from intracortical recordings in human visual cortex

Jacob Granley¹ (), Lily M. Turkstra¹, Galen Pogoncheff¹, Fabrizio Grani², Leili Soo², Alfonso Rodil², Cristina Soto², Thomas C. Sprague¹, Eduardo Fernandez², Michael Beyeler¹; ¹University of California, Santa Barbara, ²University of Miguel Hernandez, Elche, Spain

Neural activity in early visual cortex (EVC) is known to contribute to visual working memory (WM) and mental imagery (MI), but the role of spiking activity in humans remains unclear. This study investigates computational techniques for extracting spiking activity and their ability to reveal correlates of WM and MI. Intracortical recordings were collected from two awake blind humans implanted with a 96-channel visual prosthesis in EVC during a delayed-match-to-sample (DMTS) WM task and a MI visualization task. In 465 trials of the WM task, participants encoded visual perceptions (phosphenes) elicited by stimulation of one of three electrodes, maintained them over a 5-second delay, and recalled whether a subsequent phosphene was the same or different. The MI task followed a similar structure, with recall replaced by vivid mental visualization. Neural activity during stimulation, delay, recall, and spontaneous periods was analyzed using methods to extract multi-unit activity (MUA), entire spiking activity (ESA), and local field potential (LFP) signals. Significant differences were observed in MUA, ESA, and LFP (theta, alpha, and beta bands) across trial periods (t-tests, p < 0.05). ESA and MUA exhibited electrode-specific neural signatures during delay and recall periods, with over 90% classification accuracy in leave-one-trial-out cross-validation (LOOCV). Stimulus-specific ESA changes remained decodable throughout delay and recall (random forest classifier sliding window, LOOCV, 70% of windows above chance), indicating sustained stimulus-selective activity for both tasks despite day-to-day variability. These findings reveal sustained stimulus-selective spiking activity in human EVC during WM and MI tasks, underscoring its critical role in retaining and recalling information and providing new insights into the neural mechanisms underlying perception and cognition.

NIH DP2-LM014268 to MB; Alfred P. Sloan Foundation Research Fellowship to TCS; PDC2022-133952-100, PID2022-141606OB-I00 from the Spanish Ministerio de Ciencia, Innovación y Universidades, Grant No. 899287 (NeuraViPeR) from European Union’s Horizon 2020 Research and Innovation Programme to EFJ.

Talk 4, 11:30 am

How Long-Term Learning Alters Visual Working Memory Representations: Evidence from EEG

Philipp Musfeld¹, William X. Q. Ngiam², Kirsten C. S. Adam³, Olga Kozlova³, Olya Bulatova⁴, Keisuke Fukuda^4,5; ¹University of Zurich, ²The University of Adelaide, ³Rice University, ⁴University of Toronto, ⁵University of Toronto Mississauga

Visual Working memory (VWM), our central system for temporarily holding visual information in mind for further thought and action, is limited in capacity. To overcome such limitations, we frequently leverage prior knowledge from visual long-term memory (VLTM), allowing us to integrate and represent information more efficiently. Yet, it is not well understood how prior knowledge affects the representation of information in VWM. Here, we assessed how prior learning affects the load of VWM representations by using multivariate load classification from EEG. Participants (N=30) learned a 6-color visual array to criterion, and then completed a VWM task including both new and pre-learned arrays. Crucially, new arrays differed in set size (0, 1, 2 or 6), which we used to train a classifier to identify the load in VWM from the multivariate EEG signal. After establishing strong classification accuracy (~55%; chance = 25%), we asked the classifier to predict the load elicited by pre-learned arrays. We find evidence that the availability of VLTM for pre-learned arrays reduced load in VWM, as the classifier predicts a load of 1 or 2, instead of 6 – the actual set size of pre-learned arrays. However, further exploration revealed that representations of pre-learned arrays were still dissociable from pure VWM representations of lower set sizes. This was further supported by representational similarity analyses, which suggested that the obtained representational pattern for pre-learned arrays was best explained by a model assuming 1) a reduction in load in VWM together with 2) a distinct contribution from VLTM to the representation. We conclude that the availability of prior knowledge reduces the load in VWM but leads to qualitative changes in multivariate neural signals, potentially rendering memory representations more distinctive.

This research was supported by the Natural Sciences and Engineering Research Council (5009170).

Talk 5, 11:45 am

Behavioral and Neural Evidence for Dissociable Subprocesses within Visual Working Memory

Gayathri Satheesh¹ (), A. J. Abdujabborov¹, Kartik K. Sreenivasan^1,2; ¹Division of Science and Mathematics, NYU Abu Dhabi, ²Center for Brain and Health, NYU Abu Dhabi

Is working memory (WM) a unitary construct or is it composed of separable sub-processes with distinct neuroanatomical bases? In line with this latter view, WM engages a wide network of brain regions (Brissenden et al., 2018; Rahmati et al., 2020) and neurological patients display deficits in specific aspects of WM rather than global impairments (Cañas et al., 2018; Lee et al., 2010). We tested whether WM is composed of dissociable processes using a task requiring participants to maintain the locations of multiple discs over a memory delay and report the location of the cued disc. The task selectively engaged putative subcomponents of WM (storage, selection, distractor resistance, updating, and manipulation) by varying the set size, presence of task-irrelevant information during encoding or maintenance, and requirements to reformat or manipulate memory content. We collected behavioral data (n = 200) across two sessions and compared performance within and across subcomponents over sessions. Employing dimensionality reduction techniques and model fitting, we found that participants’ behavior was best described by a model that includes at least four subcomponents. These results argue against a unitary WM construct. Next, we collected fMRI data (n = 30) to determine whether these putative subcomponents map onto distinct neural substrates. We found heterogeneous activation clusters that were unique to individual subcomponents as well as regions of overlapping activation. To further tease apart the neural basis of WM subcomponents, we used representational similarity analysis to compare activity within and between subcomponents across regions of interest. The resulting dissimilarity matrix was compared to theoretical models of possible subcomponent combinations. The best-fitting models indicated distinct activity patterns across subcomponents, particularly in frontal and parietal regions. Overall, our neural data was also best explained by a multi-subcomponent model. Together, these findings indicate that WM involves multiple subcomponents with distinct patterns of neural activation.

This work was supported by the NYUAD Center for Brain and Health, funded by Tamkeen under NYUAD Research Institute grant CG012, and the ASPIRE Award for Research Excellence (AARE-19-230).

Talk 6, 12:00 pm

Neural oscillations enable concurrent visual perception and visual working memory processing

Khayla Santiago¹, Chunyue Teng¹; ¹Lawrence University, Appleton, WI

Successful goal-directed behavior requires balancing task-relevant information stored in working memory and the continuous processing of incoming sensory input, yet the neural mechanisms underlying this coordination remain unclear. The current study investigates the role of neural oscillations in supporting concurrent visual perception and visual working memory. We recorded electroencephalogram (EEG) while healthy human participants performed a dual task paradigm that requires simultaneous engagement of working memory and perceptual processing. Participants were instructed to maintain a specific orientation in mind while also observing another orientation patch on the screen. After a variable stimulus onset asynchrony (SOA) , they were prompted to compare a test probe against either the memorized orientation or the visually monitored orientation. Critically, we manipulated the duration of the SOA: 500-1500 ms with a 20 ms step, resulting in a total of 50 SOAs. Response time (RT) and accuracy were analyzed separately for each task. Visual inspection of the time courses revealed notable fluctuations in both RT and accuracy. We performed Fast Fourier transform of the data to extract spectral power and phase angle across different frequencies, and the analysis identified increased power within the theta and low-alpha frequencies for both the perceptual and memory tasks. Importantly, the two representations fluctuated at different phase angles at those identified frequencies, indicating a distinct rhythmic alternation in attentional sampling between external and internal visual representations. Additionally, Inverted Encoding Models (IEM) were applied on EEG data, and successfully reconstructed the orientations of both visual working memory and perceptual representations during periods of concurrent task relevance. Together, these results demonstrate the rhythmic nature of attentional shifts between internal and external visual representations, and further highlight the functional relevance of neural oscillations in segregating visual representations of different sources.

Talk 7, 12:15 pm

Electrical Stimulation of Visual Cortex Enhances Visual Working Memory Fidelity

Xinger Yu¹ (), Gengshi Hu¹, Geoffrey Woodman¹; ¹Vanderbilt University

Theories of visual memory suggest that the same neurons responsible for object perception also store the memory representations of these objects. This study tests the sensory recruitment hypothesis of visual working memory, which predicts that enhancing perceptual precision improves the fidelity of memory representations. We employed a combination of noninvasive transcranial direct current stimulation (tDCS), behavioral testing, and electrophysiological measures to causally manipulate neural activity in the visual cortex and evaluate its effect on memory performance. Experiment 1 involved a continuous report visual working memory task in which participants recalled the colors of items presented in varying set sizes (one, two, four, or six items). Anodal tDCS over the visual cortex (P1 or P2 site, International 10-20 System) enhanced memory precision, particularly by reducing the variability in recall errors for items located in the hemifield contralateral to the stimulation site. Experiment 2 used a change detection paradigm with retrocues, demonstrating increased accuracy for items in the contralateral hemifield following anodal tDCS, whereas no improvement was observed during sham stimulation. Simultaneous EEG recordings revealed greater alpha power suppression during the maintenance phase of the working memory tasks with anodal stimulation, suggesting a role for attention in enhancing memory performance. These findings provide causal evidence that tDCS can modulate visual cortex activity, thereby improving the fidelity of visual working memory representations. The results underscore the functional overlap between perception and memory, supporting the sensory recruitment hypothesis and highlighting the potential of noninvasive brain stimulation to improve cognitive performance.

This work was supported by grants from the National Science Foundation (BCS-2147064) and the National Eye Institute (T32-EY007135), as well as funding by the G. Forrest Woodman Foundation for the Advancement of Neuroscience.