A Visual Behavioral Training Study of Categorical Face Pattern Recognition in Mice.

A Visual Behavioral Training Study of Categorical Face Pattern Recognition in Mice - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Ann N Y Acad Sci . 2026 Jun 4;1560(1):e70308. doi: 10.1111/nyas.70308 A Visual Behavioral Training Study of Categorical Face Pattern Recognition in Mice Yanhong Chen 1 , Lingna Guo 2 , Ri Jiang 2 , Ke Zhou 3 , Xudong Zhao 2, 4, 5, ✉ , Ming Meng 1, ✉ Author information Article notes Copyright and License information 1 Department of Neurobiology, University of Alabama at Birmingham, Birmingham, Alabama, USA 2 State Key Laboratory of Cognitive Science and Mental Health, Institute of Biophysics, Chinese Academy of Sciences, Beijing, China 3 Department of Psychology, Beijing Normal University, Beijing, China 4 CAS Center for Excellence in Brain Science and Intelligence Technology, Beijing, China 5 University of Chinese Academy of Sciences, Beijing, China ✉ Corresponding author. Revised 2026 Apr 6; Received 2025 Oct 30; Accepted 2026 Apr 27; Issue date 2026 Jun. © 2026 The Author(s). Annals of the New York Academy of Sciences published by Wiley Periodicals LLC on behalf of The New York Academy of Sciences. This is an open access article under the terms of the http://creativecommons.org/licenses/by-nc-nd/4.0/ License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non‐commercial and no modifications or adaptations are made. PMC Copyright notice PMCID: PMC13238314  PMID: 42244244 ABSTRACT Can rodents form a categorical face concept or is their ability to discriminate faces from objects driven solely by visual features? To answer this question, we trained C57BL/6J mice on a touchscreen face–object discrimination task and examined their behavior in two experiments. We show that: (i) after training, mice generalized face–object classification to novel exemplars but lost the ability to discriminate when stimuli were inverted or contrast‑negated, consistent with the classic effects of these manipulations on humans performing categorically face‐specific tasks; (ii) along a continuum from low‑similarity pareidolia to real faces, choice performance followed a sigmoidal rather than a linear or quadratic function, which again suggests a categorical face concept recognition boundary. These behavioral results in mice resemble how humans visually categorize faces and nonfaces. The work provides evidence that a face concept can emerge even in a species whose social communication is not primarily visual, paving the way for future investigations on neural and genetic substrates of categorical pattern recognition. Keywords: animal behavior, concept categorization, contrast negation, face pareidolia, inversion effect, mouse vision Mice trained in a touchscreen face–object discrimination task generalized to novel exemplars but showed reduced performance when images were inverted or contrast‐negated. Along a continuum from pareidolia to real faces, choice behavior followed an S‐shaped curve, with mouse face–object discrimination showing patterns similar to those seen in humans. 1. Introduction Faces convey crucial information about survival and social interactions, necessitating that both humans and animals often detect faces swiftly and accurately. Whether faces constitute a special object category processed by dedicated mechanisms or merely by general‐purpose visual processes has long been debated extensively [ 1 , 2 , 3 , 4 ]. Many classic studies, such as the inversion effect [ 5 ], contrast negation [ 6 ], prosopagnosia [ 7 ], and focal fusiform face area activity [ 8 ], build a basis for face‑specific mechanisms. Yet proponents of domain‐general account argue that faces are handled by the same general visual processes used for all visual categories, without requiring an evolutionarily unique module, challenging a hard‑wired face‑module account [ 9 , 10 , 11 , 12 ]. In addition to abundant primate investigations, evidence from other species offers a unique perspective on the question of whether face perception may rely on domain‐specific mechanisms at an earlier evolutionary stage. Training‐ or task‐based studies show that several nonprimate species can discriminate or categorize faces or face‐like stimuli. Sheep were trained to choose rewarded flock‑mate photos over unfamiliar ones, and still made the correct choice more than 2 years later. Moreover, neurons in their temporal and prefrontal cortices remained selectively tuned to those faces they had remembered [ 13 ]. Honeybees ( Apis mellifera ) trained in a differential‐conditioning paradigm classify simple face‑like configurations against scrambled or pattern controls and generalize the classification to novel exemplars [ 14 ]. Pigeons ( Columba livia ) likewise show population‑level neural discrimination between faces and scrambled images, and neuronal recordings in their visual forebrain reliably separate five stimulus categories [ 15 ]. A separate line of comparative work has examined spontaneous orienting to face‐like stimuli before extensive visual learning. Using parallel paradigms and closely matched stimuli, studies in newly hatched chicks and human newborns have revealed similar orienting preferences for face‐like configurations [ 16 ], and related work indicates that such visual predispositions to social stimuli are present very early in human development [ 17 ]. In chicks, spontaneous preference for face‐like stimuli has been shown under visually naïve conditions [ 18 ], and this preference is impaired when the stimuli are contrast‐negated [ 19 ]. Related work further suggests that early orienting in visually naïve chicks may also be guided by an unlearned preference for the head region of naturalistic animal stimuli [ 20 ]. These findings suggest that unlearned biases toward face‐like stimuli may provide an early scaffold for later social perception [ 21 ]. However, what may underlie the animal behavior that involves face pattern recognition? On the one hand, categorical face versus nonface discrimination would imply a unique face‐processing mechanism. Evidence from the animal behavior literature supporting this possibility includes face‑specific inversion and contrast‐negation effects in nonhuman primates [ 22 , 23 ], whereas in rats these manipulations similarly reduce performance [ 24 ], and in newborn chicks contrast negation reduces spontaneous preference for face‐like stimuli [ 19 ]. On the other hand, visual classification behavior could also be achieved through learned associations between visual semblance to target exemplar features and reward, without requiring the formation of a conceptual face category. Moreover, the use of a binary stimulus set (i.e., face vs. nonface) limits interpretability, as performance may reflect simple pairwise image matching rather than a categorical representation of faces. To overcome this challenge, we utilized a graded image set containing stimuli with varying degrees of facial similarity. A central component of this continuum consists of pareidolia images [ 25 , 26 ]. The pareidolia images appear face‐like due to their spatial configuration, yet lack identity or concept of face category, offering a way to disentangle perceptual similarity from categorical representation. With this strategy, the right fusiform gyrus in humans has already been shown to display a clear categorical boundary along such a similarity continuum, which separates faces from nonfaces. In contrast, activation in the left fusiform gyrus varied more linearly with face‐likeness, consistent with a graded encoding of perceptual similarity [ 25 ]. Recent magnetoencephalogram–fMRI data further indicate that pareidolia images initially evoke neural responses indistinguishable from those elicited by real faces, followed by a rapid reclassification as nonface objects [ 27 ]. Interestingly, rhesus monkeys also exhibit robust pareidolia responses to face‑like patterns [ 28 ], suggesting perhaps a broadly defined face categorization. However, it remains unknown whether rodents such as mice possess a comparable categorical boundary representation or instead rely on the visually learned association of feature semblance with reward. Unlike primates, rodents do not rely on vision for pivotal social interactions. If rodents, too, display categorical recognition of faces, this would imply a domain‐specific face concept rather than domain‐general visual pattern recognition. In other words, if rodents could do so, the way they learn to recognize face patterns may, even more than primates, suggest both domain‐specific categorization of faces and some necessary domain‐general visual‐processing mechanisms. Thus, here we conducted two behavioral experiments to test face recognition in mice. In Experiment 1, we first show that mice can be trained to discriminate face images, including both human and mouse faces, from nonface objects on a touchscreen and to generalize this discrimination to previously unseen exemplars. Next, we explore whether the trained mice exhibit face recognition performances common to humans and other species through established image manipulations uniquely associated with face recognition processes, including experiments with contrast‐negated and inverted images. In Experiment 2, by using a graded image set containing stimuli with varying degrees of facial similarity (e.g., face pareidolia), we explore whether mice can form a categorical face concept or merely process face pareidolia stimuli through domain‐general and continuous visual feature analyses. 2. Materials and Methods 2.1. Animals We conducted experiments using 16 eight‐week‐old C57BL/6J mice, housed individually in separate cages within the research vivarium room. The mice were transported to the facility at 7 weeks old and allowed 2 days of acclimation with ad libitum access to food before their weight was recorded. Throughout the experimental phase, the body weight of the mice was maintained at 85%–95% of their initial measures. The vivarium was subjected to controlled lighting conditions (lights on from 7:00 p.m. to 7:00 a.m., and off from 7:00 a.m. to 7:00 p.m.) to ensure that the mice were energetically active during the daytime. Each experiment was conducted in the morning, with pre‐experimental weighing and postexperimental refeeding with standard chow according to weight. All animal experiments were approved by the Ethics Committee of the Institute of Biophysics, Chinese Academy of Sciences (Approval Number: SYXK2023167). 2.2. Setup The study employed a MED touchscreen system, a variant of the Skinner box housed in an anechoic chamber and linked to an external control box. The behavioral task was a simultaneous two‐choice visual discrimination task. Within the experimental chamber, a touchscreen of 1024 × 768 pixels refreshing at 60–72 Hz was divided by a central partition into two stimulus areas measuring 7.5 cm × 7.5 cm each. Mice were assigned to either a human‑face group ( n = 8) or a mouse‑face group ( n = 8), and on every trial, one face image (human or mouse, depending on group) and one object image were randomly selected and randomly presented on the left or right side of the screen. Throughout daily 30‐min training sessions, the mouse advanced through as many as 60 trials. A touch to the face image was defined as the correct response and immediately triggered delivery of a food pellet from a feeder at the rear of the chamber. A touch to the object image was defined as an incorrect response and was followed by a light cue as a no‑reward punishment. An error‑correction procedure required repetition of the same trial until the correct choice was made, with a limit of three corrections. Training continued until a mouse reached at least 70% accuracy on 2 consecutive days and was evaluated by the experimenter to be ready for formal testing. Formal testing consisted of two experiments. Experiment 1 presented novel human faces, mouse faces, and object images, together with inverted and contrast‑negated versions of other novel face images. Experiment 2, conducted only with the human‑face group after completion of Experiment 1, presented images with low, medium, or high human‑face similarity. In both experiments, each mouse completed 60 trials per day: 48 trials using training images and 12 trials using test images. As in training, each test trial consisted of a simultaneous choice between one target image and one object image. For every image set, the target image appeared once on each screen side across the 12 test trials, which were administered without reward or light cues. For each trial, response time (the interval between image onset and screen touch) and behavioral response (1  =  correct, 0  =  incorrect) were recorded. In Experiment 1, a correct response was any selection of an upright, inverted, or contrast‑negated face image over an object; in Experiment 2, it was a selection of either a real face or a face pareidolia image over an object. 2.3. Stimulus The stimuli included images of human faces, mouse faces, objects, and face pareidolia stimuli. The mouse‐face images were captured by the experimenters, while the human face, object, and face pareidolia images were sourced from a computer‐recognized image database created by Meng et al. [ 25 ]. This image database comprised 300 images, including 60 nonface images, 60 human‐face images, and 180 face pareidolia images. The face pareidolia images were derived as algorithmic false positives using the Pittsburgh Pattern Recognition face‐detection algorithm from the aforementioned sets. Furthermore, in the original study by Meng et al. [ 25 ], 18 participants performed a forced‐choice task on image pairs, selecting the image that more closely resembled a face in each trial. The images were subsequently re‐rated using the Elo rating algorithm. As a result, the image database includes both human selection proportions and Elo ratings. Our face pareidolia images were categorized into low, medium, and high facial similarity, with computer and human ratings closely aligned within each similarity category. All images were converted to grayscale and matched for luminance, contrast, and spatial frequency using the SHINE Toolbox [ 29 ]. Moreover, we conducted image‐similarity analyses of the training and test stimuli. Pairwise similarity between the face and object images within each training or test pair was quantified using histogram of oriented gradients (HOG) cosine similarity and an edge‐based Dice coefficient. The similarity between each test face and the upright face images used during training was assessed separately using the structural similarity index measure (SSIM). Detailed information can be found in the Supporting Information . In addition, we recruited 20 participants to provide human‐based face similarity ratings using a 10‐point Likert scale ranging from “very dissimilar to a human face” to “identical to a real human face.” This human rating procedure was approved by the Human Research Ethics Committee for Non‐Clinical Faculties, School of Psychology, South China Normal University (Reference No.: SCNU‐PSY‐2023‐380), and all participants provided written informed consent before participation. 2.4. Statistical Analyses All analyses were conducted in R‐4.2.3 [ 30 ]. 2.4.1. Training For evaluating the training performance, we charted each mouse's learning curve by modeling daily accuracy (defined as the percentage of selecting faces) with generalized additive models (GAMs). Accuracy percentages were converted to numeric values, and separate smooths were fit for every mouse to capture nonlinear trends. We overlaid a 70% criterion line and flagged the final two sessions per animal to summarize asymptotic performance. 2.4.2. Experiment 1 Initially, we sought to determine whether mice could (i) generalize the learned face–object discrimination to novel faces and (ii) reveal the inversion and contrast‐negation effects. Thus, we constructed generalized linear mixed‐effects models (GLMMs) with the following formula for mouse performance from the human‐face and mouse‐face groups separately: M 1 : Response ∼ Condition − 1 + 1 | ID + 1 | Image + 1 | Date where condition, which is the fixed effect term, represents generalization (new face images that have not been shown in the training phase), the inversion condition, and the contrast‐negation condition. −1 removes the intercept term from the model, which allows each condition to be compared against a 50% chance level. The response variable represented the choices made by the mice: correct choices were encoded as 1, whereas incorrect choices were encoded as 0. The models also included random intercepts for individual mouse IDs, the specific images used in the recognition task, and the date of testing. 2.4.3. Experiment 2 In Experiment 2, we continued testing with the mice from the human‐face group. We categorized the experimental materials into three categories based on facial similarity: low, medium, and high. Each category included two sets of images, each with consistent facial‐similarity ratings as assessed by both computer algorithms and human participants. We included the generalization condition from Experiment 1 as the real‐face condition and combined it with three different facial‐similarity conditions. We then established a GLMM with model terms similar to M1: M 2 : Response ∼ Condition − 1 + 1 | ID + 1 | Image + 1 | Date For comparing accuracy (defined as selecting face or face pareidolia) between the real‐face condition and the other three different facial‐similarity conditions, we established the following model, with the real‐face condition as a reference: M 3 : Response ∼ Condition + 1 | ID + 1 | Image + 1 | Date We aimed to determine whether the mice's responses to images with varying levels of facial similarity were based on categorical or linear changes. If the mice responded to these differences in a categorical manner, we would expect their response to follow an S‐shaped curve. To investigate this, we fitted the facial‑similarity data and the corresponding recognition success rates with sigmoid, quadratic, and linear functions. Because the mice were exposed to a small set of distinct test images, the dataset available for fitting the sigmoid model remained correspondingly sparse. To overcome this limitation, we augmented the data by simulating additional data points. Specifically, we utilized the human ratings for the anthropomorphism of the images to calculate the mean scores and standard deviations (SD) for each image. Similarly, we calculated the mean response and SD for each image based on the mice's responses. We then generated additional data points for each image using bivariate Gaussian distributions, where the mean vector was composed of the mean probability of correct responses and the mean rating, and the covariance matrix was constructed using the SD and a correlation coefficient to account for the relationship between these two variables. For each simulation, we generated 1000 data points per image, creating a larger dataset for analysis. This simulated dataset was then used to fit regression models based on sigmoid, quadratic, and linear functions. The sigmoid function was fitted using a nonlinear least squares (NLS) approach with the following model: M 4 : ACC ∼ L 1 + exp − k × Rating − x 0 where L represents the upper asymptote, k is the slope parameter of the curve, and x 0 is the inflection point or the Rating value at which the fitted response reaches half of the upper asymptote ( L /2). Initial estimates for these parameters were specified, with x 0 starting at the median of the Rating values. To evaluate the fit of the sigmoid model, we calculated an R 2 value. For comparison, we also fitted a linear model: M 5 : ACC ∼ Rating and a quadratic model: M 6 : ACC ∼ poly Rating , 2 The quadratic model, with its additional parameter, provided a more balanced comparison with the sigmoid function, which also contains three parameters, and served as a control to ensure that any improvement in fit for the sigmoid function was not merely due to the increased complexity of the model. To assess the robustness of our findings, we conducted 2000 simulations. We calculated 95% confidence intervals for the pairwise differences in R 2 values to determine which of the sigmoid, linear, or quadratic equations yielded the best overall fit. 3. Results 3.1. Training Due to poor learning performance, one mouse from each group was excluded (Mouse 3 from the human‐face group and Mouse 4 from the mouse‐face group, as shown in Figure 1 ). FIGURE 1. Open in a new tab Learning curves for each mouse. The first and second rows show the learning curves of mice in the human‐face group ( N = 8), while the third and fourth rows show those of mice in the mouse‐face group ( N = 8). The red dashed line indicates the 70% accuracy threshold, and the last two red dots represent performance on the final 2 days of training. The x ‐axis for each subplot represents the training date; the y ‐axis represents the averaged face‐selection accuracy for each day. 3.2. Experiment 1 In two GLMMs, by comparing task accuracy between picture types and chance levels for human‐ and mouse‐face training groups, we found that mice are able to discriminate between new face pictures and objects (Human face: Mean ACC = 0.604, p = 2.25 × 10 −5 ; Mouse face: Mean ACC = 0.559, p = 0.009). By contrast, performances on inverted (Human face: Mean ACC = 0.537, p = 0.132; Mouse face: Mean ACC = 0.532, p = 0.152) and contrast‐negated (Human face: Mean ACC = 0.504, p = 0.884; Mouse face: Mean ACC = 0.535, p = 0.124) images were not significantly different from chance levels (Figure 2A,B ). FIGURE 2. Open in a new tab Bar plots for the performance of the human‐ and mouse‐face groups under three conditions. (A) GLMM for the human‐face group. (B) GLMM for the mouse‐face group. Red dashed lines indicate the chance level, y ‐axes represent the predicted probability of selecting face images for each stimulus condition by GLMM. Error bars indicate 95% confidence intervals (CIs) We also analyzed the response time data from Experiment 1. The results indicated that there were no significant main effects for group or condition, nor were there significant interactions observed. 3.3. Experiment 2 The accuracy results indicated that the probability of selecting real human‐face images significantly exceeded the chance level (Mean ACC = 0.606, p = 4.76 × 10 −4 ). By contrast, accuracy for low‑, medium‑, and high‑similarity images did not differ from chance, although the high‑similarity condition showed a marginal trend toward significance (Mean ACC  =  0.558, p =  0.0628). For the model setting real human faces as the reference group, accuracy results (Figure 3A ) showed significant differences between low facial similarity and real faces ( Z = −2.822, p = 0.005), and medium facial similarity and real faces ( Z = −2.845, p = 0.004), but no significant difference between high facial similarity and real faces ( Z = −1.144, p = 0.253). Analysis based on response times (Figure 3B ) indicated significant differences between the low facial similarity ( t = 3.184, p = 0.009), medium facial similarity ( t = 3.748, p = 0.004), and high facial similarity ( t = 2.773, p = 0.020) with the real‐face group. FIGURE 3. Open in a new tab GLMMs for performances on different face similarity data. (A) The GLMM that sets the real human‐face group as a reference. The y ‐axes show the predicted probability of selecting face images for each stimulus condition by GLMM, and the red dashed lines indicate the chance level. (B) The GLMM that sets the real human‐face group as a reference. The y ‐axis represents the predicted LogRT for each face stimulus condition by GLMM. In both panels, the x ‑axis, NF L , NF M, and NF H refer to pareidolia faces of low, medium, and high similarity to real faces, respectively. Error bars indicate 95% CIs. The highest explanatory ability was observed in the sigmoid function (Mean = 0.119, SD = ±0.005), followed by the quadratic model (Mean = 0.112, SD = ±0.005), the lowest was the linear model (Mean = 0.103, SD = ±0.005); differences between the sigmoid function and the linear (95% CI [0.011, 0.022]) and quadratic (95% CI [0.004, 0.011]) model were significant. Because this model comparison was based on simulation‐augmented data, the apparent advantage of the sigmoid model should be interpreted cautiously as preliminary evidence for a sigmoidal tendency. 4. Discussion With two experiments, our study convergently suggests that face concepts may emerge in mice. We utilized a Skinner box to train and test the face recognition capabilities of mice. After extensive training, mice demonstrated the ability to distinguish between faces and objects, and such ability was generalized to new exemplars. However, their performance declined when presented with inverted and contrast‐negated images, consistent with the inversion and contrast‐negation effect of face‐specific processing in humans. Furthermore, when mice were presented with images varying in facial similarity, they categorized these images in a manner similar to humans, as indicated by the highest explanatory ability of a sigmoid function. Such a sigmoidal response indicates the presence of a categorical boundary, suggesting that these mice perceived a categorical division between “face” and “nonface” stimuli along the continuum. While performance generalized to other upright and positive exemplars, the loss of discrimination with inverted and contrast‐negated images suggests that the performance of mice depended on the orientation and natural contrast structure of the learned stimuli, properties that are also known to be very important for human‐face recognition [ 31 ]. For humans, inverting a face image disrupts its configural layout and impairs recognition far more than inverting other objects, a phenomenon regarded as a behavioral marker of face‐specific processing [ 5 , 32 , 33 ]. Here in our study, mice displayed a comparable vulnerability: upright faces were readily recognized as “face,” but upside‐down faces were not. In addition, face‐detection models have emphasized the importance of correct contrast polarity relationships between facial regions for identifying faces [ 36 ]. Human and nonhuman primate studies convergently show that contrast negation dramatically hinders face recognition by disrupting these luminance‐based cues [ 23 , 37 , 38 ]. Likewise, mice could not generalize visual discrimination to contrast‐negated faces, indicating that they also relied on prototypical light–dark polarity patterns for face processing. However, this result alone does not determine whether the mice relied on the holistic configuration of features, the outer contour, local facial features, or other image‐level properties for the discrimination task. We performed two image analyses addressing two possible strategies. First, if the mice solved the task by comparing the two alternatives presented together on each trial and choosing on the basis of their relative visual difference, then the poorer performance for inverted and contrast‐negated stimuli should have been accompanied by a substantial increase in pairwise similarity between the face and object images. However, the pairwise similarity of the training and test pairs, quantified by HOG cosine similarity and an edge‐based Dice coefficient, was broadly comparable and did not show a consistent increase for the inversion or contrast‐negation conditions. This suggests that the impairment for inverted and contrast‐negated stimuli is unlikely to be fully explained by simple within‐trial image similarity or increased pairwise confusability. Second, if the mice judged each test face according to how much it resembled the upright face images encountered during training, then performance might instead depend on similarity to the trained positive exemplars. Consistent with this possibility, an SSIM‐based image‐similarity classifier performed well for the upright generalization condition but poorly for the inversion and contrast‐negation conditions, suggesting that the transformed faces differed substantially from the learned upright face images. Nevertheless, these two analyses cannot identify the specific cues driving behavior. To determine whether mice relied on configural information per se rather than broader image‐level similarity, future studies will need to use more selective manipulations, such as isolating internal features while preserving contour or selectively disrupting second‐order relations [ 34 , 35 ]. The face versus nonface categorization boundary in Experiment 2 further implies the emergence of face concept in mice. Under the classical framework of statistical decisions, projecting onto a discriminant yields posterior class probabilities that change nonlinearly, often captured by logistic/cumulative‑normal functions, near the decision boundary [ 39 , 40 ]. By contrast, exemplar/nearest‑neighbor style matching predicts a broadly monotonic (not abrupt) increase in accuracy with similarity [ 41 ]. We thus tested whether the mice's psychometric function exhibits a steep, boundary‑like transition indicative of categorical perception [ 42 , 43 ] or a smoother monotonic trend consistent with similarity matching of visual features. The results of Experiment 2 suggest that a sigmoid (S‐shaped) function fits the behavioral performance of mice across varying degrees of face similarity better than a linear or even quadratic function. In other words, as stimuli became more face‐like, the choice behavior of mice showed an abrupt transition from chance‐level to above‐chance classification of faces, rather than a gradual, linear improvement. This nonlinear performance pattern is a hallmark of categorical perception [ 42 ], indicating that the mice may have formed a face versus nonface categorization boundary. The finding that the sigmoid model provided the best fit was derived from simulated data, but these simulated data were constrained by the observed image‐level central tendency, variability, and covariation of human face‐likeness ratings and mice's face‐choice probabilities, such that the analysis remained anchored to the empirical structure of the original data. In this sense, the simulation provides an informative way to examine whether a boundary‐like pattern would remain plausible under data generated from the observed distributional properties. At the same time, because the augmented dataset was generated from the original measurements rather than collected as new observations, the simulation should be understood as a way of probing the stability of the observed pattern under empirically constrained conditions, rather than as an independent source of evidence. Its value therefore lies in showing that a boundary‐like interpretation remains plausible when the data are augmented in a manner consistent with the observed distributional structure, but it is not sufficient on its own to establish a definitive face versus nonface categorization boundary. Accordingly, the apparent sigmoidal pattern is best interpreted as supportive rather than definitive evidence for such a boundary. Future studies using larger and more diverse stimulus sets, together with denser sampling across levels of face‐likeness, will be important for determining whether this apparent sigmoidal relationship reflects a robust property of mouse‐face categorization. Note that although the average accuracy for high‑similarity pareidolia images was statistically indistinguishable from that for real faces, all three similarity levels of pareidolia images elicited significantly longer reaction times. According to the classic bounded‑accumulation model [ 44 ], the decision variable reflects the moment‑to‑moment accumulation of “faceness” evidence until it crosses a categorical threshold. A real face therefore should supply strong evidence, quickly reaching the criterion. High‑similarity pareidolia however, would supply slightly weaker evidence: it may eventually cross the same threshold, yielding a face‑level choice, but the weaker signal would prolong accumulation time, producing a reliable RT cost. Since the present task was a face/pareidolia versus object discrimination task, the rate of this accumulation would also depend on how readily each face‐like stimulus could be distinguished from the competing object alternative. This speed‐accuracy asymmetry resembles what has been well documented in human research for any stimuli near category boundaries [ 45 ]. While our study may suggest visual categorization in mice, this boundary does not have to overlap with the same boundary for humans, whose visual system has adapted to discriminate more precisely between genuine faces and nonfaces. How rodents and primates may visually categorize faces versus nonfaces with the same or different exact boundaries awaits further research. On the one hand, the inversion effect, the contrast‐negation effect, and the categorical boundary in mice convergently suggest the engagement of a specialized face‐processing mechanism, much as in primates. At the same time, mice required extensive training and reward feedback before they could reliably perform the face/nonface discrimination. A conservative interpretation, therefore, is that repeated exposure, reinforcement, and task demands can progressively tune a flexible visual system toward explicit face/nonface categorization, rather than that mice spontaneously express a mature face specialization under baseline conditions. This interpretation is consistent with developmental and comparative evidence. In monkeys, inferotemporal representations can be described within a low‐dimensional object space, with face‐selective regions embedded in a broader coarse map of object space [ 48 ]. Monkeys deprived of face exposure from birth fail to develop these face‐selective domains despite otherwise normal vision [ 49 ], and human‐face specialization appears to be substantially refined through postnatal experience [ 50 ]. On the other hand, face perception likely reflects contributions from both experience‐dependent learning and early face‐oriented predispositions, especially for face detection [ 51 ]. Consistent with this possibility, previous studies in human newborns and newly hatched, visually naïve chicks have documented spontaneous orienting toward face‐like configurations, including effects of contrast negation and, in chicks, an unlearned preference for the head region, suggesting early unlearned biases toward socially relevant patterns [ 16 , 17 , 18 , 19 , 20 , 55 ]. Recent neurophysiological studies further indicate cortical responses to face‐like patterns in human newborns and face‐selective neurons in face‐naïve young chicks [ 46 , 47 ]. Related, though not directly face‐selective, comparative work in visually naïve chicks has linked early responses to naturalistic social stimuli to differential activity in the intermediate medial mesopallium and the nucleus taeniae of the amygdala [ 53 , 54 ]. Some deprivation paradigms in monkeys have also reported an initial behavioral preference for faces even in the absence of normal face exposure [ 56 ], although the emergence of canonical face‐selective cortical domains still appears to depend on face experience [ 49 ]. Our study shows that mice can acquire a face category through training and that the learned behavior exhibits several hallmarks commonly associated with face processing. At the same time, our results do not rule out the possibility that broader unlearned visual biases may provide an initial scaffold for later face learning. These results demonstrate that categorical face recognition, once considered exclusive to primates, can also be acquired by mice. A critical next step is to identify the underlying neural representation. In primates, discrete cortical regions and single neurons respond preferentially to faces [ 8 , 52 ]. It remains unknown whether comparable units emerge in the mouse visual system after training. Future studies could employ techniques like two‐photon calcium imaging or electrophysiological recordings in visual areas to identify neurons or circuits that preferentially respond to face images over nonfaces. If such face‐selective signals are confirmed, follow‑up work should determine whether their tuning reflects configural, category‑level coding rather than sensitivity to isolated visual features. So far, rodents are not known for strong innate face preferences, so any face concept in mice is likely a product of experience‐dependent plasticity, making the mouse a useful model for studying how experience constructs a new visual category. Genetic and circuit‑level tools can isolate the molecules and pathways that support this learning and can test whether selective disruption of low‑level visual processing impedes the emergence of the face concept. These approaches will help disentangle the contributions of general visual competence and category‑specific mechanisms. 5. Conclusion In summary, our study finds that mice may be trained to form a face category instead of merely responding to incremental visual similarity. Our findings provide novel evidence in a nonprimate species of some core behavioral manifestations of the formation of domain‐specific face concepts. This not only broadens our understanding of the evolutionary roots of face recognition, but also establishes the feasibility of using animal models to study the neural and genetic basis of high‐level visual categorization. Author Contributions Yanhong Chen: conceptualization, methodology, formal analysis, data curation, visualization, writing – original draft, writing – review and editing. Lingna Guo: methodology (experimental program setup), investigation (behavioral training/testing), data curation (data categorization), resources (animal husbandry/maintenance; mouse photographs). Ri Jiang: resources (animal husbandry/maintenance; mouse photographs). Ke Zhou: conceptualization, methodology (advice on image‐processing procedures). Xudong Zhao: conceptualization, methodology, data curation (data hosting/organizational guidance), supervision, writing – review and editing. Ming Meng: conceptualization, funding acquisition, project administration, resources, supervision, writing – original draft, writing – review and editing. Declaration of AI Use We have not used AI‐assisted technologies in creating this article. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Supporting Information : nyas70308‐sup‐0001‐SuppMat.docx NYAS-1560-0-s001.docx (22.1KB, docx) Acknowledgments This study was supported by grants from the STI2030‐Major Projects (2021ZD0204200), the Sino‐German Center for Research Promotion (M‐0705), the National Natural Science Foundation of China (32371100), the Research Center for Brain Cognition and Human Development, Guangdong, China (No. 2024B0303390003), and the Striving for the First‐Class, Improving Weak Links and Highlighting Features (SIH) Key Discipline for Psychology in South China Normal University. We acknowledge Shanghai Youer Inc. and Chao Li for technical assistance in the programming of the K‐limbic software. Contributor Information Xudong Zhao, Email: zhaoxudong@ibp.ac.cn. Ming Meng, Email: mengm@uab.edu. References 1. Cohen J. D. and Tong F., “The Face of Controversy,” Science 293 (2001): 2405–2407, 10.1126/science.1066018. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Kanwisher N., “Domain Specificity in Face Perception,” Nature Neuroscience 3 (2000): 759–763, 10.1038/77664. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Haxby J. V., Gobbini M. I., Furey M. L., Ishai A., Schouten J. L., and Pietrini P., “Distributed and Overlapping Representations of Faces and Objects in Ventral Temporal Cortex,” Science 293 (2001): 2425–2430, 10.1126/science.1063736. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Duchaine B. and Yovel G., “A Revised Neural Framework for Face Processing,” Annual Review of Vision Science 1 (2015): 393–416, 10.1146/annurev-vision-082114-035518. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Yin R. K., “Looking at Upside‐Down Faces,” Journal of Experimental Psychology 81 (1969): 141–145, 10.1037/h0027474. [ DOI ] [ Google Scholar ] 6. Kemp R., Pike G., White P., and Musselman A., “Perception and Recognition of Normal and Negative Faces: The Role of Shape From Shading and Pigmentation Cues,” Perception 25 (1996): 37–52, 10.1068/p250037. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Damasio A. R., Damasio H., and Van Hoesen G. W., “Prosopagnosia: Anatomic Basis and Behavioral Mechanisms,” Neurology 32 (1982): 331–341, 10.1212/wnl.32.4.331. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Kanwisher N., McDermott J., and Chun M. M., “The Fusiform Face Area: A Module in Human Extrastriate Cortex Specialized for Face Perception,” Journal of Neuroscience 17 (1997): 4302–4311, 10.1523/JNEUROSCI.17-11-04302.1997. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Diamond R. and Carey S., “Why Faces Are and Are Not Special: An Effect of Expertise,” Journal of Experimental Psychology: General 115 (1986): 107–117, 10.1037//00963445.115.2.107. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Gauthier I., Tarr M., Anderson A., Skudlarski P., and Gore J., “Activation of the Middle Fusiform ‘Face Area’ Increases With Expertise in Recognizing Novel Objects,” Nature Neuroscience 2 (1999): 568–573, 10.1038/9224. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Gauthier I. and Tarr M. J., “Becoming a “Greeble” Expert: Exploring Mechanisms for Face Recognition,” Vision Research 37 (1997): 1673–1682, 10.1016/S0042-6989(96)00286-6. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Bukach C. M., Gauthier I., and Tarr M. J., “Beyond Faces and Modularity: The Power of an Expertise Framework,” Trends in Cognitive Sciences 10 (2006): 159–166, 10.1016/j.tics.2006.02.004. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Kendrick K. M., da Costa A. P., Leigh A. E., Hinton M. R., and Peirce J. W., “Sheep Don't Forget a Face,” Nature 414 (2001): 165–166, 10.1038/35102669. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Avarguès‐Weber A., Portelli G., Benard J., Dyer A., and Giurfa M., “Configural Processing Enables Discrimination and Categorization of Face‐Like Stimuli in Honeybees,” Journal of Experimental Biology 213 (2010): 593–601, 10.1242/jeb.039263. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Clark W., Chilcott M., Azizi A., Pusch R., Perry K., and Colombo M., “Neurons in the Pigeon Visual Network Discriminate Between Faces, Scrambled Faces, and Sine Grating Images,” Scientific Reports 12 (2022): 589, 10.1038/s41598-021-04559-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Rosa Salva O., Farroni T., Regolin L., Vallortigara G., and Johnson M. H., “The Evolution of Social Orienting: Evidence From Chicks ( Gallus gallus ) and Human Newborns,” PLoS ONE 6 (2011): e18802, 10.1371/journal.pone.0018802. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Di Giorgio E., Frasnelli E., Rosa Salva O., et al., “Difference in Visual Social Predispositions Between Newborns at Low‐ and High‐Risk for Autism,” Scientific Reports 6 (2016): 26395, 10.1038/srep26395. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Rosa‐Salva O., Regolin L., and Vallortigara G., “Faces Are Special for Newly Hatched Chicks: Evidence for Inborn Domain‐Specific Mechanisms Underlying Spontaneous Preferences for Face‐Like Stimuli,” Developmental Science 13 (2010): 565–577, 10.1111/j.1467-7687.2009.00914.x. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Rosa Salva O., Regolin L., and Vallortigara G., “Inversion of Contrast Polarity Abolishes Spontaneous Preferences for Face‐Like Stimuli in Newborn Chicks,” Behavioural Brain Research 228 (2012): 133–143, 10.1016/j.bbr.2011.11.025. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Rosa‐Salva O., Mayer U., and Vallortigara G., “Unlearned Visual Preferences for the Head Region in Domestic Chicks,” PLoS ONE 14 (2019): e0222079, 10.1371/journal.pone.0222079. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Rosa Salva O., Mayer U., and Vallortigara G., “Roots of a Social Brain: Developmental Models of Emerging Animacy‐Detection Mechanisms,” Neuroscience & Biobehavioral Reviews 50 (2015): 150–168, 10.1016/j.neubiorev.2014.12.015. [ DOI ] [ PubMed ] [ Google Scholar ] 22. Neiworth J. J., Hassett J. M., and Sylvester C. J., “Face Processing in Humans and New World Monkeys: The Influence of Experiential and Ecological Factors,” Animal Cognition 10 (2007): 125–134, 10.1007/s10071-006-0045-4. [ DOI ] [ PubMed ] [ Google Scholar ] 23. Parr L. A. and Taubert J., “The Importance of Surface‐Based Cues for Face Discrimination in Non‐Human Primates,” Proceedings of the Royal Society B: Biological Sciences 278 (2011): 1964–1972, 10.1098/rspb.2010.2333. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Schnell A. E., Van den Bergh G., Vermaercke B., Gijbels K., Bossens C., and de Beeck H. O., “Face Categorization and Behavioral Templates in Rats,” Journal of Vision 19 (2019): 9, 10.1167/19.14.9. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Meng M., Cherian T., Singal G., and Sinha P., “Lateralization of Face Processing in the Human Brain,” Proceedings of the Royal Society B: Biological Sciences 279 (2012): 2052–2061, 10.1098/rspb.2011.1784. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Moulson M. C., Balas B., Nelson C., and Sinha P., “EEG Correlates of Categorical and Graded Face Perception,” Neuropsychologia 49 (2011): 3847–3853, 10.1016/j.neuropsychologia.2011.09.046. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Wardle S. G., Taubert J., Teichmann L., and Baker C. I., “Rapid and Dynamic Processing of Face Pareidolia in the Human Brain,” Nature Communications 11 (2020): 4518, 10.1038/s41467-020-18325-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Taubert J., Wardle S. G., Flessert M., Leopold D. A., and Ungerleider L. G., “Face Pareidolia in the Rhesus Monkey,” Current Biology 27 (2017): 2505–2509.e2, 10.1016/j.cub.2017.06.075. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Willenbockel V., Sadr J., Fiset D., Horne G. O., Gosselin F., and Tanaka J. W., “Controlling Low‐Level Image Properties: The SHINE Toolbox,” Behavior Research Methods 42 (2010): 671–684, 10.3758/BRM.42.3.671. [ DOI ] [ PubMed ] [ Google Scholar ] 30. R Core Team R . R: A Language and Environment for Statistical Computing (Citeseer, 2020). [ Google Scholar ] 31. Sinha P., Balas B., Ostrovsky Y., and Russell R., “Face Recognition by Humans: Nineteen Results all Computer Vision Researchers Should Know About,” Proceedings of the IEEE 94 (2006): 1948–1962, 10.1109/JPROC.2006.884093. [ DOI ] [ Google Scholar ] 32. Yovel G. and Kanwisher N., “The Neural Basis of the Behavioral Face‐Inversion Effect,” Current Biology 15 (2005): 2256–2262, 10.1016/j.cub.2005.10.072. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Valentine T., “Upside‐Down Faces: A Review of the Effect of Inversion Upon Face Recognition,” British Journal of Psychology 79 (1988): 471–491, 10.1111/j.2044-8295.1988.tb02747.x. [ DOI ] [ PubMed ] [ Google Scholar ] 34. Mondloch C. J., Le Grand R., and Maurer D., “Configural Face Processing Develops More Slowly Than Featural Face Processing,” Perception 31 (2002): 553–566, 10.1068/p3339. [ DOI ] [ PubMed ] [ Google Scholar ] 35. Rotshtein P., Geng J. J., Driver J., and Dolan R. J., “Role of Features and Second‐Order Spatial Relations in Face Discrimination, Face Recognition, and Individual Face Skills: Behavioral and Functional Magnetic Resonance Imaging Data,” Journal of Cognitive Neuroscience 19 (2007): 1435–1452, 10.1162/jocn.2007.19.9.1435. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Sinha P., “Qualitative Representations for Recognition,” in Biologically Motivated Computer Vision (LNCS 2525), ed. Bülthoff H. H., Wallraven C., Lee S. W., and Poggio T. A. (Springer, 2002), 249–262, 10.1007/3-540-36181-2_25. [ DOI ] [ Google Scholar ] 37. Pallett P. M. and Meng M., “Contrast Negation Differentiates Visual Pathways Underlying Dynamic and Invariant Facial Processing,” Journal of Vision 13 (2013): 13, 10.1167/13.14.13. [ DOI ] [ PubMed ] [ Google Scholar ] 38. Ohayon S., Freiwald W. A., and Tsao D. Y., “What Makes a Cell Face Selective? The Importance of Contrast,” Neuron 74 (2012): 567–581, 10.1016/j.neuron.2012.03.024. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Fisher R. A., “The Use of Multiple Measurements in Taxonomic Problems,” Annals of Eugenics 7 (1936): 179–188, 10.1111/j.1469-1809.1936.tb02137.x. [ DOI ] [ Google Scholar ] 40. Cox D. R., “The Regression Analysis of Binary Sequences,” Journal of the Royal Statistical Society Series B: Statistical Methodology 20 (1958): 215–232, 10.1111/j.2517-6161.1958.tb00292.x. [ DOI ] [ Google Scholar ] 41. Cover T. and Hart P., “Nearest Neighbor Pattern Classification,” IEEE Transactions on Information Theory 13 (1967): 21–27, 10.1109/TIT.1967.1053964. [ DOI ] [ Google Scholar ] 42. Harnad S. R., ed., Categorical Perception: The Groundwork of Cognition (Cambridge University Press, 1987). [ Google Scholar ] 43. Ashby F. G. and Maddox W. T., “Human Category Learning,” Annual Review of Psychology 56 (2005): 149–178, 10.1146/annurev.psych.56.091103.070217. [ DOI ] [ PubMed ] [ Google Scholar ] 44. Gold J. I. and Shadlen M. N., “The Neural Basis of Decision Making,” Annual Review of Neuroscience 30 (2007): 535–574, 10.1146/annurev.neuro.29.051605.113038. [ DOI ] [ PubMed ] [ Google Scholar ] 45. Ratcliff R. and McKoon G., “The Diffusion Decision Model: Theory and Data for Two‐Choice Decision Tasks,” Neural Computation 20 (2008): 873–922, 10.1162/neco.2008.12-06-420. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Buiatti M., Di Giorgio E., Piazza M., et al., “Cortical Route for Facelike Pattern Processing in Human Newborns,” Proceedings of the National Academy of Sciences 116 (2019): 4625–4630, 10.1073/pnas.1812419116. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Kobylkov D., Rosa‐Salva O., Zanon M., and Vallortigara G., “Innate Face‐Selectivity in the Brain of Young Domestic Chicks,” Proceedings of the National Academy of Sciences 121 (2024): e2410404121, 10.1073/pnas.2410404121. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Bao P., She L., McGill M., and Tsao D. Y., “A Map of Object Space in Primate Inferotemporal Cortex,” Nature 583 (2020): 103–108, 10.1038/s41586-020-2350-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Arcaro M. J., Schade P. F., Vincent J. L., Ponce C. R., and Livingstone M. S., “Seeing Faces Is Necessary for Face‐Domain Formation,” Nature Neuroscience 20 (2017): 1404–1412, 10.1038/nn.4635. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Scott L. S. and Arcaro M. J., “A Domain‐Relevant Framework for the Development of Face Processing,” Nature Reviews Psychology 2 (2023): 183–195, 10.1038/s44159-023-00152-5. [ DOI ] [ Google Scholar ] 51. Kobylkov D. and Vallortigara G., “Face Detection Mechanisms: Nature Vs. Nurture,” Frontiers in Neuroscience 18 (2024): 1404174, 10.3389/fnins.2024.1404174. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Tsao D. Y., Freiwald W. A., Tootell R. B. H., and Livingstone M. S., “A Cortical Region Consisting Entirely of Face‐Selective Cells,” Science 311 (2006): 670–674, 10.1126/science.1119983. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Mayer U., Rosa‐Salva O., Lorenzi E., and Vallortigara G., “Social Predisposition Dependent Neuronal Activity in the Intermediate Medial Mesopallium of Domestic Chicks ( Gallus gallus domesticus ),” Behavioural Brain Research 310 (2016): 93–102, 10.1016/j.bbr.2016.05.019. [ DOI ] [ PubMed ] [ Google Scholar ] 54. Mayer U., Rosa‐Salva O., Loveland J. L., and Vallortigara G., “Selective Response of the Nucleus Taeniae of the Amygdala to a Naturalistic Social Stimulus in Visually Naive Domestic Chicks,” Scientific Reports 9 (2019): 9849, 10.1038/s41598-019-46322-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Farroni T., Johnson M. H., Menon E., Zulian L., Faraguna D., and Csibra G., “Newborns' Preference for Face‐Relevant Stimuli: Effects of Contrast Polarity,” Proceedings of the National Academy of Sciences 102 (2005): 17245–17250, 10.1073/pnas.0502205102. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Sugita Y., “Face Perception in Monkeys Reared With no Exposure to Faces,” Proceedings of the National Academy of Sciences 105 (2008): 394–398, 10.1073/pnas.0706079105. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supporting Information : nyas70308‐sup‐0001‐SuppMat.docx NYAS-1560-0-s001.docx (22.1KB, docx) Articles from Annals of the New York Academy of Sciences are provided here courtesy of Wiley ACTIONS View on publisher site PDF (869.0 KB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• FreqFLD: Towards All-in-One Facial Landmark Detection via Frequency Modulation#997175
• Early Abstraction, Continuous Generalization: The Dynamics of Real-Life Face Learning#663374
• Early Abstraction, Continuous Generalization: The Dynamics of Real-Life Face Learning#663248
• Cerebral correlates and statistical criteria of cross-modal face and voice integration#183691
• Cerebral correlates and statistical criteria of cross-modal face and voice integration#183541