Face and Body Perception: Development, clinical, individual differences, experience

Talk Session: Tuesday, May 20, 2025, 5:30 – 7:15 pm, Talk Room 2

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

Talk 1, 5:30 pm

An age bias in other-race face perception: neural and behavioral evidence

Moaz Shoura¹, Yong Zhong Liang¹, Dirk B. Walther¹, Adrian Nestor¹; ¹University of Toronto

The other-race effect (ORE) refers to a well-documented disadvantage in recognizing faces of other races compared to one’s own. Yet, the perceptual biases underlying ORE remain less understood. This research combines behavioral similarity measures, neural decoding and style-based generative adversarial networks (StyleGAN2; Karras et al., 2020) to examine visual biases, with focus on age misrepresentation in other-race (OR) face perception. To this end, first, East Asian and White participants (n = 190) rated the pairwise visual similarity of OR and same-race (SR) synthetically generated faces. The similarity structure of behavioral data was mapped onto that of StyleGAN2 latent representations. Relying on this mapping, we employed a novel procedure for GAN-based image reconstruction to recover SR and OR face percepts across our participants. The procedure generated hyper-realistic visualizations of face percepts. More importantly, facial attribute analysis, using RetinaFace (Deng et al., 2019), revealed that OR image reconstructions appeared systematically younger than SR ones. Second, electroencephalography (EEG) data were collected from East Asian and White participants (n = 40) who viewed OR and SR faces. Neural representations of these faces, recovered through EEG-based face decoding and image reconstruction (Nemrodov et al., 2018), were subsequently rated by another group of East Asian and White validators (n = 46) for perceived age and typicality. OR face reconstructions were consistently rated as younger and more typical for their race than SR reconstructions across both validator groups. Together, these findings provide convergent evidence for age misrepresentation in OR face perception across different data types and different reconstruction methodologies. These results open new avenues for investigating representational differences in face perception and highlight new implications for cross-racial interactions.

Talk 2, 5:45 pm

Category selectivity observed in the human FFA, PPA, and EBA is distinct from category selectivity observed in artificial neural network units

Alish Dipani^1,2 (), N Apurva Ratan Murty^1,2; ¹Cognition and Brain Science, School of Psychology, Georgia Institute of Technology, ²Center of Excellence in Computational Cognition, Georgia Institute of Technology

Category-selective responses to faces, scenes, and bodies have been robustly observed in the human brain (e.g. FFA, PPA, and EBA) and more recently, even in artificial neural network (ANN) models. This raises a key question: How similar are the category-selective responses observed in the human brain compared to the category-selective responses seen in vision ANNs? To address this question, we first identified the category-selective voxels and ANN units following an identical localization procedure. Consistent with prior findings, we successfully isolated the category-selective voxels and ANN model units. Next, we compared the voxel-averaged response patterns in the FFA, PPA, and EBA (from the Natural Scenes Dataset) with the unit-averaged response patterns observed in category-selective ANN units. Response patterns observed in the brain were remarkably consistent across participants (median subject-subject R = 0.68). In contrast, the response patterns of category-selective ANN units were significantly distinct from those observed in category-selective voxels (median subject-ANN R = 0.38). This divergence between ANN units and brain regions was robust: it persisted for all ANN models evaluated (N=13) and across different localizers (N=3), generalized to multivariate tests between model units and brains, and held across all brain regions tested. Finally, we asked whether category-selective ANN units could be linearly combined (encoding model) to predict responses in human category-selective regions. Surprisingly, while the encoding models had high predictivity, lesioning the category-selective ANN units did not significantly impact the model’s prediction accuracy. This finding is further evidence that category-selective ANN units may not directly correspond to human category-selective brain regions. Together, these results underscore the need for caution: not all forms of category selectivity can be considered equivalent. The category selectivity observed in FFA, PPA, and EBA is a specific kind, fundamentally distinct from the category selectivity exhibited in ANN units.

This work is supported by a CoCo fellowship by the School of Psychology, Georgia Tech to AD and by the NIH Pathway to Independence Award (R00EY032603), and a startup grant from Georgia Tech to NARM.

Talk 3, 6:00 pm

From Eyes to Identity: Early Neural Sensitivity to the Eye Region is Associated with Face Recognition Ability

Anthony Proulx¹ (), Isabelle Charbonneau¹, Justin Duncan², Vicki Ledrou-Paquet¹, Chanelle Demeule¹, Caroline Blais¹, Daniel Fiset¹; ¹Université du Québec en Outaouais, ²Ottawa University

Recognizing faces is a fundamental social skill that varies widely among individuals. While most people identify faces effortlessly, others experience significant difficulties, with extreme cases characterized as prosopagnosia. Research on individual differences and prosopagnosia has emphasized the importance of the eye region in face identification (Caldara et al., 2005; Tardif et al., 2019), with greater reliance on the eyes predicting better face recognition ability (Royer et al., 2018). However, this association may arise from different underlying mechanisms. It is possible that the superior use of the eyes results from more efficient perceptual processing, better facial memory, or a combination of both. To explore these hypotheses, 36 participants completed a series of face and object recognition tasks, with a principal component analysis (PCA) generating a global face recognition ability score. Participants then performed a face identification task while their brain activity was recorded using EEG. Faces were presented through small Gaussian apertures (“Bubbles”; Gosselin & Schyns, 2001), and classification images were generated from -200ms to 800ms post-stimulus, revealing pixels associated with stronger or weaker voltage at PO8. These images were aligned with each participant’s N170 and N250 latencies and combined to capture feature processing within these time windows. Weighted averages based on PCA scores highlighted individual differences in classification images relative to face recognition ability. Pixel tests (p<.05; Stat4Ci Toolbox; Chauvin et al., 2005) on these weighted images revealed that better face recognition ability was significantly associated with increased sensitivity to both the contra- and ipsilateral eyes shortly before the N170, and with increased sensitivity to the contralateral eye before the N250. These results suggest that better face recognizers make superior use of the eye region early in perceptual processing, which also manifests as more effective utilization of the contralateral eye when activating memory representations.

Talk 4, 6:15 pm

Predicting individualized category-selective functional topographies in developmental prosopagnosia using connectivity hyperalignment

Ian Abenes¹ (), Jiahui Guo; ¹The University of Texas at Dallas

Category-selective functional topographies are mostly similar across individuals, but considerable variability exists in the exact topographical location, size, and shape of these areas. It has been demonstrated that individualized category-selective topographies can be estimated with high fidelity using hyperalignment (Jiahui et al., 2020, Jiahui et al., 2023). However, previous work only included typical participants, and it is unclear whether this method can be extended to neuropsychological populations. To address this, we analyzed data from 12 individuals with developmental prosopagnosia (DP), who display profound face recognition deficits in the absence of brain lesions or broad neurodevelopmental problems, alongside 16 typical participants. We first estimated functional topographies of four categories: faces, bodies, objects, and scenes from a classic dynamic localizer. We then used connectivity hyperalignment (CHA) to derive transformation matrices for each individual using a separate functional scan from other participants within their group. This scan was taken while participants completed a one-back task to identify if the previously shown face was the same in identity, expression, or view. We applied the transformations back to the localizer data to project all other participants’ localizer data to the given individual’s space. The estimated topographies using CHA for the given individual were the mean across all other participants’ contrast maps. Finally, we estimated the topographies using anatomical alignment (AA) by directly averaging contrast maps based on surface-aligned localizer runs within each group. We correlated each participant's own localizer map with the CHA or AA estimated maps to measure the performance of the predictions. On average, CHA showed higher correlations than AA for estimating individual face-selective topography, and similar results were found for other categories. Notably, both DP and typical participants showed similar high-fidelity estimations, suggesting successful estimation of category-selective topographies using connectivity hyperalignment, despite significant impairments in face recognition abilities in the DP individuals.

Talk 5, 6:30 pm

Category selectivity as an explanation for multidimensional tuning in human occipitotemporal cortex

Hans Op de Beeck¹, Elahe' Yargholi¹; ¹KU Leuven, Leuven Brain Institute, Dpt. Brain & Cognition, Leuven, Belgium

Recent studies have revealed an exceptionally rich landscape of functional selectivity in human occipitotemporal cortex. While the first seminal brain imaging studies of object recognition emphasized the existence of category-selective regions, alternative perspectives have emerged. Explicit comparisons suggest that dimensions outperform categories in predicting brain responses (Contier et al., 2024, Nature Human Behavior), and that the tuning of category-selective neurons can be explained by domain-general tuning for nonface dimensions (Vinken et al., 2023, Science Advances). This begs the question whether the notion of category selectivity is outdated and should be abandoned in favor of a dimensional view (Ritchie et al., 2024, arXiv). Here we review existing findings and present new findings that illustrate the power of the notion of category selectivity to explain object representations and the observed dimensional tuning profiles. Previous studies suggest that the most important dimensions in the literature, animacy and stubby/spiky, could possibly be attributed to category selectivity. Furthermore, in a new fMRI study (N = 22) with participants viewing social scenes that all contain the same categories, we find a strong modulation of neural responses in face-, body-, and scene-selective regions by variation in how categories are depicted in natural images, consistent with realistic category-based models. Overall, the notion of category selectivity provides a parsimonious explanation for the strongest forms of selectivity which are observed in human occipitotemporal cortex.

METH/24/003

Talk 6, 6:45 pm

A new computational model of human face recognition that learns continuously by generating images from memory

Naphtali Abudarham¹, Galit Yovel¹; ¹Tel Aviv University

In recent years, deep convolutional neural networks (DCNNs) have reached human-level performance in face recognition and have been shown to exhibit human-like face effects including the face inversion effect, the other race effect, and sensitivity to human-like critical features. Nevertheless, there are still significant discrepancies between the operations of DCNNs and the human face system. One fundamental discrepancy is the training regime – DCNNs are trained in batch mode, learning the whole training set at once, while humans learn to recognize different identities gradually over time. Another discrepancy is that humans can learn and recognize new faces instantly after a single exposure, whereas training DCNNs on new identities requires extensive exposure to each identity and may lead to catastrophic forgetting of previously learned identities. To bridge these gaps, we propose a novel computational cognitive model for human face recognition. Our model is composed of three main components: 1. An embedder, which is trained incrementally on a small number of faces in each step, using a Continual Learning paradigm and is used to create face representations. 2. A face-generator module, which is trained in parallel to the embedder and is used to generate images of old faces for training the embedder concurrently with new faces. 3. A memory module, which stores embedding statistics of familiar/trained identities, enabling the recognition of familiar faces using a nearest-neighbor search, and instant learning of new identities. We show how this model can learn and recognize faces nearly as good as a batch model, even when it is trained on a very small number of identities. Furthermore, our model shows a human-like familiarity benefit on a face sorting task. Our model proposes a comprehensive and realistic approach to human face recognition, which can be expanded beyond faces to study mechanisms of human memory.

Talk 7, 7:00 pm

Examining the functional organization of marmoset inferotemporal cortex using calcium imaging

David G. C. Hildebrand¹, Santiago Otero-Coronel¹, Alipasha Vaziri¹, Winrich Freiwald¹; ¹Rockefeller University

A characteristic feature of sensory cortex in primates is its functional organization into continuous maps, where neurons are arranged according to their functional properties (e.g., retinotopy, orientation pinwheels). In contrast, high-level visual areas in inferotemporal (IT) cortex have been parcellated into discrete category-selective clusters (e.g., face vs. body areas). However, recent results and models suggest that IT cortex could instead be functionally organized by continuous maps corresponding to properties such as animacy or real-world size. It has been difficult to resolve the functional organization of IT cortex with standard techniques; fMRI has insufficient spatial resolution, while electrophysiological sampling is too sparse. To examine the functional organization of IT cortex, we developed an approach for cellular-resolution imaging of neuronal populations in marmosets. The marmoset cortex is smooth, making it accessible with optical imaging. Leveraging this feature, we localized face areas using intrinsic signal imaging. We then recorded calcium dynamics of cortical layer 2/3 neurons in awake, head-restrained marmosets using two-photon microscopy with fields of view up to 3×3 mm2 while presenting faces, non-face objects, and bodies. This approach allowed us to sample thousands of IT cortical neurons in and around face areas. While neurons in face areas responded more on average to faces, the stimulus eliciting the peak response from each neuron revealed a striking diversity across the population. Nearby neurons tended to respond more to similar subsets of face and non-face stimuli. Furthermore, we observed gradients of selectivity between face, object, or body areas rather than discrete boundaries, supporting the conclusion that IT cortex is functionally organized in a continuous map.

NIH R21 EY031486