About

Professor and Interim Head of the Department of Psychology at the University of Illinois, Diane M. Beck, PhD, is a principal investigator at the Attention and Perception Lab within the Beckman Institute. Her research focuses on attention and perception, contributing to the understanding of cognitive processes related to visual attention and perceptual mechanisms. She collaborates with a network of researchers and has a history of mentoring graduate students and postdoctoral scholars, advancing the field through her leadership and scholarly work.

Research topics

• Computer Science
• Artificial Intelligence
• Mathematics
• Physics
• Cognitive psychology
• Nuclear physics
• Psychology
• Natural Language Processing
• Algorithm
• Materials science
• Mathematical analysis
• Linguistics
• Chemistry
• Atomic physics
• Statistics
• Optics
• Quantum mechanics
• Particle physics

Selected publications

• Dissociating spatial frequency reliance from adversarial robustness advantages in neurally guided deep convolutional neural networks

  PubMed Central · 2026-05-06

  preprintOpen accessSenior author

  Deep convolutional neural networks (DCNNs) have rivaled humans on many visual tasks, yet they remain vulnerable to near-imperceptible perturbations generated by adversarial attacks. Recent work shows that aligning DCNN representations with human visual cortex activity improves adversarial robustness, but the mechanisms driving this advantage are unclear. One hypothesis suggests that neural alignment confers robustness by biasing models away from brittle high-frequency details and towards the low spatial frequencies (LSF). However, recent work shows that human object recognition critically depends on a narrow, mid-frequency "human channel". Interestingly, this band was partially preserved in prior LSF-focused studies. Here, we investigate whether a spectral bias towards the LSF or the human channel is the primary driver of the adversarial robustness observed in neurally aligned DCNNs. We first show that DCNNs aligned to higher-order regions of the human ventral visual stream systematically increase reliance on both LSF and the human channel. However, directly steering DCNNs towards these bands revealed a clear dissociation. Biasing models towards the human channel, either alone or together with LSF, does not improve robustness and even impairs it. LSF bias produced some robustness gains, but such improvements are modest despite inducing much larger shifts in spatial-frequency reliance than neurally aligned models. Spatial-frequency-biased models overall show little, if any, increase in similarity to human neural representational geometry. Together, our results suggest that altered spatial-frequency reliance is likely an emergent property of learning more human-like representations rather than the primary mechanism by which neural alignment confers adversarial robustness, and motivate the need for future research examining representational properties beyond spatial-frequency profiles.

  PublisherDOI

• Dissociating spatial frequency reliance from adversarial robustness advantages in neurally guided deep convolutional neural networks

  ArXiv.org · 2026-05-06

  articleOpen accessSenior author

  Deep convolutional neural networks (DCNNs) have rivaled humans on many visual tasks, yet they remain vulnerable to near-imperceptible perturbations generated by adversarial attacks. Recent work shows that aligning DCNN representations with human visual cortex activity improves adversarial robustness, but the mechanisms driving this advantage are unclear. One hypothesis suggests that neural alignment confers robustness by biasing models away from brittle high-frequency details and towards the low spatial frequencies (LSF). However, recent work shows that human object recognition critically depends on a narrow, mid-frequency "human channel". Interestingly, this band was partially preserved in prior LSF-focused studies. Here, we investigate whether a spectral bias towards the LSF or the human channel is the primary driver of the adversarial robustness observed in neurally aligned DCNNs. We first show that DCNNs aligned to higher-order regions of the human ventral visual stream systematically increase reliance on both LSF and the human channel. However, directly steering DCNNs towards these bands revealed a clear dissociation. Biasing models towards the human channel, either alone or together with LSF, does not improve robustness and even impairs it. LSF bias produced some robustness gains, but such improvements are modest despite inducing much larger shifts in spatial-frequency reliance than neurally aligned models. Spatial-frequency-biased models overall show little, if any, increase in similarity to human neural representational geometry. Together, our results suggest that altered spatial-frequency reliance is likely an emergent property of learning more human-like representations rather than the primary mechanism by which neural alignment confers adversarial robustness, and motivate the need for future research examining representational properties beyond spatial-frequency profiles.

  PublisherOA PDF

• The Neural and Behavioral Correlates of Music Memorability

  2026-05-09

  articleOpen accessSenior author
  PublisherOA PDFDOI

• Dissociating spatial frequency reliance from adversarial robustness advantages in neurally guided deep convolutional neural networks.

  PubMed · 2026-05-06

  articleSenior author

  Deep convolutional neural networks (DCNNs) have rivaled humans on many visual tasks, yet they remain vulnerable to near-imperceptible perturbations generated by adversarial attacks. Recent work shows that aligning DCNN representations with human visual cortex activity improves adversarial robustness, but the mechanisms driving this advantage are unclear. One hypothesis suggests that neural alignment confers robustness by biasing models away from brittle high-frequency details and towards the low spatial frequencies (LSF). However, recent work shows that human object recognition critically depends on a narrow, mid-frequency "human channel". Interestingly, this band was partially preserved in prior LSF-focused studies. Here, we investigate whether a spectral bias towards the LSF or the human channel is the primary driver of the adversarial robustness observed in neurally aligned DCNNs. We first show that DCNNs aligned to higher-order regions of the human ventral visual stream systematically increase reliance on both LSF and the human channel. However, directly steering DCNNs towards these bands revealed a clear dissociation. Biasing models towards the human channel, either alone or together with LSF, does not improve robustness and even impairs it. LSF bias produced some robustness gains, but such improvements are modest despite inducing much larger shifts in spatial-frequency reliance than neurally aligned models. Spatial-frequency-biased models overall show little, if any, increase in similarity to human neural representational geometry. Together, our results suggest that altered spatial-frequency reliance is likely an emergent property of learning more human-like representations rather than the primary mechanism by which neural alignment confers adversarial robustness, and motivate the need for future research examining representational properties beyond spatial-frequency profiles.

  Publisher

• What makes good exemplars of a scene category good? Evidence from deep neural nets

  Vision Research · 2025-11-17

  articleSenior author
  PublisherDOI

• Human Visual Robustness Emerges from Manifold Disentanglement in the Ventral Visual Stream

  Journal of Vision · 2025-07-15

  articleOpen accessSenior author

  Humans effortlessly navigate the dynamic visual world, yet deep neural networks (DNNs), despite excelling in visual tasks, are surprisingly vulnerable to image perturbations that are innocuous to humans. Aligning DNN representations with human neural representations, particularly those from higher-order regions of the ventral visual stream (VVS), has been shown to improve their robustness (Shao et al., 2024). Such observation suggests that the representational space in the VVS has desirable properties that support human robustness but are absent in DNNs. One particular framework posits that the VVS achieves robust inference by progressively disentangling neural category manifolds (Dicarlo & Cox, 2007). Specifically, neural population responses to identity-preserving changes of objects form continuous manifolds in the neural state space. These manifolds are initially tangled, i.e., linearly inseparable, but become progressively disentangled across stages of the VVS, naturally resulting in robust inference. Despite its theoretical appeal, empirical evidence for this framework remains limited. Here, using a computational characterization of neural manifold statistics (Chung et al., 2018) and a 7T fMRI dataset (Allen et al., 2022), we first demonstrate that category manifolds at different stages of the human VVS show increasingly desirable geometric properties: smaller radius and compressed dimensionality, together leading to improved overall linear separability. Importantly, we show that these properties are inheritable by DNNs through neural representation alignment and indeed predict subsequent robustness gains observed in previous work. Finally, to more directly test this framework, we propose “manifold guidance”, a method that aligns DNNs to human VVS on the granularity of category manifolds, without imposing strict individual representation matching commonly adopted in previous neural alignment studies. We show that manifold guidance is capable of leading to robustness improvements in DNNs. Our findings, thus, provide compelling evidence that human visual robustness arises from the disentanglement of category manifolds in the VVS.

  PublisherDOI

• Three categorization tasks yield comparable category spaces: a comparison using real-world scenes

  Journal of Vision · 2025-07-15

  articleOpen accessSenior author

  Categorization is a fundamental cognitive function that depends on the similarity among items. Many tasks have been developed to measure item similarities to better understand the underlying similarity space of categories. Most such studies have focused on object categories. The current study compares the measured similarity outcomes of three categorization tasks for real-world scenes: an arrangement, odd-one-out, and same-different judgment task. Our study asks whether a stable representation for scene categories (i.e., beach, city, mountain) exists across tasks. To assess the reliability of the tasks, each task was conducted twice for each participant (N = 98 for each task). The arrangement task asked participants to place each scene image relative to the three text anchors that they set at the beginning of the task. The odd-one-out task required participants to choose the scene that is the most dissimilar to the other two in a triplet of scenes. The same-different judgment task asked participants to respond whether a pair of scenes was from the same category or not. The similarity matrices were derived from distances in pixel space for the arrangement task, the probability of choosing a scene as dissimilar for the odd-one-out task, and the probability of reporting two scenes as ‘same’ for the same/different task. The rank correlations were calculated between two repeats of the task to examine the reliability of the similarities. All three tasks showed comparable rank correlation reliability: arrangement (0.61), odd-one-out (0.60), same-different (0.69). Ordinal multidimensional scaling on the similarity matrices of each task was used to construct 3-D category spaces, where the distances reflect the similarity among stimuli. These rank-derived spaces were moderately correlated across tasks: arrangement & same-different (0.67), arrangement & odd-one-out (0.58), odd-one-out & same-different (0.57). These results imply some stable representation of scene category similarity space that is worth further investigation.

  PublisherDOI

• Real-world Statistical Regularity in Binocular Rivalry: the advantage of good exemplars

  Journal of Vision · 2025-07-15

  articleOpen accessSenior author

  Real-world statistical regularities, unlike regularities introduced in experimental settings, are learned through extensive exposure over a lifetime in the natural visual environment. These natural patterns reflect the consistent features and structure of real-world scenes and objects. By utilizing these predictable patterns, the visual system processes statistically regular stimuli more quickly and effortlessly than less regular ones, enhancing perceptual efficiency (Beck, Center, Shao, 2024). One example of real-world statistical regularity is the distinction between good and bad exemplars images. Good exemplars, which are highly representative of a category, are more easily detected than bad exemplars. Building on this, the current study aimed to investigate whether statistical regularity influences perceptual selection in binocular rivalry, where conflicting images presented to each eye compete for the dominance of perception. In this study, participants were shown two images from the same scene category (i.e., beach, mountain, city, highway). One image was a good exemplar, and the other was a bad exemplar of that category, with each image presented to a different eye. Results revealed statistical regularity biased the perceptual selection to the good exemplar: good exemplars were more likely to be selected as the initial percept and had faster perceptual onset times compared to bad exemplars. Our results align with the predictive coding framework of binocular rivalry, where good exemplars, with higher priors, would be more likely to dominate perception over bad exemplars. These findings extend findings on statistical learning (e.g. Denison, Piazza, Silver, 2011) to real-world statistical regularities that are derived over a lifetime rather than within an experiment.

  PublisherDOI

• Effects of task type on spontaneous alternations of attentional states

  Memory & Cognition · 2024-01-18 · 1 citations

  articleOpen access
  PublisherOA PDFDOI

• Image memorability is linked to facilitated perceptual and semantic processing

  Imaging Neuroscience · 2024-01-01 · 8 citations

  articleOpen access

  Strikingly, some images are consistently more likely to be remembered compared to others-a stable, intrinsic image property that has been termed image memorability. However, the properties that afford this memory advantage have remained elusive. In prior work, we showed that more memorable images are easier to perceive, and modeling work further suggests that semantic properties contribute to variance in memorability. Thus, we hypothesize that image memorability effects arise at the interface between perception and semantic memory. To test this hypothesis, we used event-related potentials (ERPs) to measure perceptual template matching (N300) and semantic access (N400) processes in a continuous recognition memory task using high and low memorability images, each repeated once. On initial presentation, both N300 and N400 amplitudes were less negative for high memorability images, showing that memorability is linked to both facilitated high-level perceptual processing and more efficient semantic activation. High memorability images also elicited a larger N300 repetition effect compared to low memorability images, revealing that their perceptual processing benefits more from the prior exposure. The results support the idea that images that better match visual templates and elicit more targeted semantic activations are easier to identify when encountered again, and further point to a potential interplay between semantic activation and perceptual matching in supporting image memorability.

  PublisherDOI

Recent grants

• NIH Grant R03MH082012

  NIH · $145k · 2009

• NIH Grant R01EY022605

  NIH · $1.3M · 2018

• A Program of Medium Energy Nuclear Physics

  NSF · $7.2M · 2015–2019

Frequent coauthors

• Li Fei-Fei

  53 shared

• W. Turchinetz

  Massachusetts Institute of Technology

  36 shared

• K. Stephenson

  Schlumberger (United States)

  36 shared

• R. Redwine

  36 shared

• E.J. Stephenson

  Indiana University Bloomington

  36 shared

• S. Kowalski

  36 shared

• R. Goloskie

  Princeton University

  36 shared

• S. Gilad

  Massachusetts Institute of Technology

  36 shared