Vision Research

Mounting evidence suggests that recurrent processes in the visual system play a critical role during challenging recognition tasks. Backward masking techniques have traditionally been used as a non-invasive method for studying recurrent processes: A mask follows the target image, presumably disrupting ongoing processes. However, these techniques have the limitation that they do not allow the identification of the stage of the visual system at which critical recurrent processes are taking place. Here, leveraging advances in texture synthesis via deep networks, and the approximate correspondence between stages of the visual system and layers of deep networks, we develop a novel psychophysics paradigm where masks with textures targeting different stages of the visual system follow the presentation of challenging images. In a series of experiments, we present objects to human subjects either for a short duration or in unusual poses, followed by a textured mask either designed to only target the early visual system, or the entire visual system. We find that both texture types equally affect recognition abilities, suggesting that recurrent processes in or towards early stages of the visual system are already recruited for these recognition tasks.

Numerosity adaptation (the underestimation of number after exposure to a numerous adaptor) is reduced when adaptor and test differ in color, suggesting that the numerosity system parses items into color-defined categories. Here we ask whether this chromatic selectivity is organized into multiple narrowly tuned chromatic channels, and whether its expression depends on individual chromatic sensitivity. Twenty observers (aged 22-61) completed two psychophysical tasks. First, chromatic discrimination was measured for five hues spaced in 5{degrees} CIE L*a*b* steps ({Delta}H = 0{degrees}, 5{degrees}, 10{degrees}, 15{degrees}, 20{degrees}) from a red reference (LCh: 54, 118, 38), yielding an individual just-noticeable difference (JND). Second, numerosity adaptation was measured across the same five chromatic distances between a 48-dot adaptor and the test. Observers with superior discrimination (JND < 2.5{degrees}) showed robust chromatic tuning, adaptation declining as the test moved away from the adaptor hue, whereas poorer discriminators showed none. Using an interindividual-covariance / factor-analytic approach, we found that adaptation strengths at neighboring chromatic distances were highly correlated and fell off with chromatic separation. Principal component analysis extracted two factors, one loading on the larger chromatic distances and one on the smaller; under oblique (promax) rotation the two factors were substantially correlated (r = .66), implying at least two dissociable but overlapping chromatically tuned mechanisms. These results suggest that numerosity adaptation is mediated by multiple, comparatively narrow chromatic channels, resembling the higher-order color mechanisms inferred from color scaling, SSVEP, and fMRI, rather than the two early cardinal axes (L-M, S-(L+M)).

Mental representations are the explanatory construct of the cognitive sciences, but there is no widely accepted characterization of how they cause behavior. Visual representational geometry can be quantified by similarity scores, but almost all methods require an explicit judgment. To examine how representations cause behavior by varying task demands, observers must perform different tasks while continuously reporting similarity, leading to dual-task interference. This study develops scanpaths as an implicit similarity measure, and uses representational similarity analysis to validate it. Observers searched for a target; fixations on distractors may reveal similarity. Similarity was also quantified by an odd-one-out task in the same participants, and ratings from different participants (Jiang et al. 2022). Representational geometries between tasks correlated. A generative model predicted first fixations in novel data. This double validation of the scanpath method opens the door to examining the causality of representations by determining if and how they vary with task demands.

Optical coherence tomography (OCT) is widely used for retinal disease assessment, but automated quantitative analysis remains challenging because of anatomical variability and noisy imaging conditions. This study presents an interpretable OCT classification framework based on four anatomically guided retinal layers, combining preprocessing, adaptive segmentation, targeted feature engineering, and supervised classification to identify Normal, CNV, DME, and Drusen cases. Layer-specific descriptors included statistical, derivative, fluid-related, and GLCM texture markers. Feature correlation and ranking analyses showed that the proposed descriptors were highly complementary, that the most informative features were concentrated in layers 2 and 4, and that layers 1 and 3 contributed supportive structural information. Among the evaluated classifiers, the neural network performed best, achieving an accuracy of 98.17%, sensitivity of 97.88%, specificity of 99.38%, and AUC of 0.9985. Computational analysis showed efficient training and inference, with a total training time of 2216.3 s, prediction speed of approximately 160000 observations per second, and a compact model size of about 11 kB. These results demonstrated that anatomically guided feature extraction can provide accurate, efficient, and interpretable OCT disease classification, offering a practical alternative to less transparent end-to-end deep learning approaches.

In this paper, we address the problem of quantifying similarity between planar 2D shapes, which is relevant to studies of internal representations in cognitive, developmental, and neurological research. We designed a set of test shapes arranged along a visually defined perceptual similarity gradient and used them to evaluate classical geometric methods for shape comparison, including Procrustes and Chamfer distance, as well as a convolutional neural network (CNN)-inspired feature-based method. Based on the limitations identified for these individual methods, we developed a hybrid Geometric-Feature Similarity (GFS) algorithm that combines geometric alignment, global contour properties, and convolutional feature-based descriptors into a unified weighted similarity score. By combining global geometric information with local structural features, the GFS algorithm more accurately reproduces human perceptual judgments of shape similarity than either geometric or feature-based methods alone. Requiring neither network training nor large labelled datasets, the proposed algorithm provides an efficient and interpretable tool for a broad range of studies involving quantitative shape comparison.

Filling-in occurs during the perceptual disappearance of a blank figure presented on a textured background. Current models of perceptual filling-in are based on a two-stage model where the figure boundary weakens after a period of adaptation, followed by the spreading of the background representation into the region representing the figure. This suggests a competition between figure boundary and background representations whereby filling-in is facilitated by a weaker boundary representation and a stronger background representation. Here, we test this interpretation, by using the oblique effect and surround-modulation suppression, which are functional properties of early visual cortex that modulate the expected strengths of the responses to the background texture and to the figure boundary. In a sample of N=58 participants, we found more filling-in with background textures of cardinal compared to oblique orientations (earlier onset time, with more and longer episodes of filling-in per trial), in line with a known, stronger neuronal response for cardinal than for oblique orientation in early visual cortex. We found more filling-in when the main axis of the rectangular figure was iso-oriented rather than cross-oriented with the background texture (more and longer episodes of filling-in per trial, but no change in onset time), in line with a lower response to oriented stimuli when surrounded by iso-oriented flankers compared to cross-oriented flankers. Overall, our results support the two-stage model and suggest the involvement of early visual cortical areas characterized by the oblique effect and orientation- tuned surround-suppression.

Introduction: Decision making for selecting an object or a course of action from possible alternatives largely depends on our perceptual ability modulated by attention. When multiple stimuli appear close together in time, processing one stimulus can temporarily impair the processing of another due to temporal limitations of attention. Observers frequently fail to detect the second target (T2) presented within a few hundred milliseconds after the first target (T1) in a stream of stimuli, which is commonly known as attentional blink (AB). Existing theories attribute this perceptual lapse to T1 processing, distractor interference, or transient attentional gating; however, the computations underlying suppressive mechanism remains unresolved. We investigated whether pupil-size could reveal the underlying mechanisms of AB and predict conscious perception on a trial-by-trial basis. Methods: Pupil diameter and gaze locations were recorded using an infrared eye tracker. Machine learning techniques were used to classify trials when T2 was detected versus when it was not, after correct identification of T1, during an AB task from the pupil dynamics, which also yielded attentional episode (AE) associated with each element in the stream of visual stimuli when deconvolved. Results: Cross-validating classifiers achieved near-perfect accuracy not only in distinguishing but also predicting perceptual outcomes on a single-trial basis. AEs exhibited greater power when T2 was detected than when it was missed; the differential power in AEs on a logarithmic scale was highly synced with the differential pupil size. Conclusions: Collectively, these findings establish a framework for predicting attention-driven perceptual outcomes from pupil-dynamics at finer time-scale.

Valproic acid (VPA) is a widely used antiepileptic drug that also increases the risk of neurodevelopmental disorders in the offspring of exposed mothers. Prenatal exposure to VPA is a widely used rodent model of autism spectrum disorder (ASD). Anatomical, functional and molecular alterations in the retinas of various ASD model animals have been described in the literature, but the impact on the neural composition of the retina remains unclear. We examined whether and how the density and spatial regularity of selected retinal neurons are altered in the VPA induced model of ASD. Whole-mount retinas of 2-month-old VPA-treated and control animals were immunolabeled for S-cones, horizontal cells, AII amacrine cells, and parvalbumin-positive wide-field amacrines (PV-wfACs), and the positions of labelled cells mapped in various regions of interest (n = 39 for treated, n = 32 for control animals) across the retinas. Multivariate analysis of variance revealed a significant overall effect of VPA on cell densities (p = 6.1x10-7, 2 = 0.43), driven mainly by reduced AII amacrine density, while horizontal cells showed a modest reduction and S-cones were unaffected. After adjusting for retinal location, analysis of covariance indicated a 7% decrease in AII cells and a 15% increase in PV-wfACs. Regularity indices calculated from nearest neighbor distances or Voronoi-domain areas of cell mosaics were largely unchanged. These findings suggest that prenatal VPA exposure selectively alters inhibitory inner retinal circuitry in the rat ASD model at the time of cell differentiation, but self-organizing mechanisms responsible for spatial order are not affected. Lay SummaryValproic acid (VPA) is a medicine for epilepsy, but it can also raise the risk of autism in children when taken during pregnancy. In rats exposed to VPA before birth, we found changes in certain nerve cells of the retina: one type of cell important for night vision was reduced, while another type increased slightly, while most other cells stayed the same. This suggests that the changes in development that lead to autism may also be reflected in the structure and function of the eye.

Fly compound eyes pool signals from photoreceptors that sample the same region of visual space through neural superposition in the lamina. The optical axes of photoreceptors projecting to a single lamina cartridge are not perfectly parallel, but instead converge at a point a few millimeters in front of the eye. At short viewing distances (1-10 mm) this leads to distance-dependent differences in receptive field overlap. We explored whether it was possible that flies could sense depth in this "personal space" purely from the geometry of neural superposition. To this end, we combined a computational model of fly eye optics with a disparity-tuned lamina model originally developed for stereoscopic prey capture in praying mantises, and simulated the responses of lamina monopolar cells to moving stimuli at different distances. Across variations in stimulus parameters and lamina models, we found that lamina neuron responses indeed contain a distance-dependent component, which can overall be summarised as an enhanced response due to temporally overlapping receptor responses at a critical distance of 3-4 mm, the convergence distance of photoreceptor axes. Depending on the lamina model and stimulus size, either response amplitude or onset gradient, or both, exhibited this peak. We further show that changes in eye size systematically shift this preferred distance, such that larger flies had a peak response at a greater distance. Our results demonstrate that lamina cell responses may contain a robust, geometry-derived component that is specific to object distance and invariant to other stimulus properties. This suggests that neural superposition, beyond improving sensitivity, may function analogously to a light-field camera system that is effectively "focused" on a behaviorally relevant distance. Author summaryIn this study, we explored how neural superposition in fly compound eyes may lead to an enhanced response to objects at a behaviorally relevant distance. The optical axis of neurally pooled photoreceptors, from neighbouring ommatidia, are not parallel, but converge at a point a few millimeters in front of the eye, which should lead to the strongest correlation of their signals at that distance. This idea was tested by combining a geometric model of the fly eye optics with a computational model of how the lamina cells process the responses of photoreceptors, and evaluating the output for a moving bright dot at different distances. We found that the simulated lamina cell has its fastest response, as measured by the onset gradient, to a stimulus at the distance where the photoreceptors converge (3-4 mm). This distance scales with the simulated eye size and corresponds to behaviorally relevant distances for fly behavior. This suggests that neural superposition may act to enhance the response to objects at a critical distance in the early visual processing of flies.

Visual discomfort (VD) is influenced by both spatial structure and chromatic context. Striped patterns are well-established triggers of discomfort and autonomic responses. In previous work, we showed that higher spatial frequencies and larger patterned areas elicit stronger pupillary constriction and greater discomfort, and that individuals with higher overall discomfort show shallower maximum constriction. The present study examined whether similar relationships appear when spatial structure is held constant, measured background luminance is kept within a narrow range, and the chromatic background varies. Participants viewed black horizontal stripes on 12 near-isoluminant colored backgrounds. The CIE76 color difference ({Delta}E) ranged from 36 to 112 and was indexed as the CIELAB distance from the black stripe pattern. Pupil size was continuously recorded and discomfort ratings were collected after each trial. Across the colored backgrounds, more uncomfortable stimuli evoked stronger pupil constriction, even though luminance was held nearly constant. As in our spatial-frequency study, this stimulus-level increase did not translate into stronger constriction among observers reporting higher overall discomfort: participants who rated the stimuli as more uncomfortable overall showed shallower constriction. This pattern was captured by the maximum-constriction response, which differentiated high-from low-discomfort observers and was significantly associated with individual discomfort ratings. Together with our previous findings, the results suggest that pupil responses scale along the tested stimulus axis, whereas individuals reporting greater visual discomfort exhibit less pronounced maximum constriction. HighlightsO_LIChromatic background modulated discomfort and pupil responses at similar luminance. C_LIO_LIAcross backgrounds, higher discomfort ratings tracked stronger pupil constriction. C_LIO_LIAcross observers, higher mean discomfort tracked weaker pupil constriction. C_LIO_LIThis two-level dissociation recurs across spatial and chromatic manipulations. C_LIO_LIPupillometry may complement subjective reports of visual discomfort. C_LI

Recurrent processing involves feedforward, feedback and lateral connections and is thought to allow efficient visual processing. Anatomical and behavioral studies in humans have suggested that feedback connections mature later in development than feedforward connections and thus were proposed to depend to a larger degree on experience. In order to isolate feedforward from feedback activity and to investigate the role of early visual experience, we assessed seven individuals with reversed congenital cataracts and nine sighted controls using an "occlusion paradigm" with 7T magnetic resonance imaging (Smith & Muckli, 2010): Grayscale images of scenes were presented with the lower right quadrant covered by a white rectangle. We examined whether information about category (beaches, buildings, highways) and individual scenes could be extracted from early visual region vertices (V1 - V3) associated with the occluded quadrant of the visual field, in the absence of bottom-up visual input. This was achieved by decoding individual category or scene context utilizing a linear support vector machine. In addition, bidirectional information flow was assessed using connective field modeling. While both groups showed successful decoding of scene and category from vertices receiving bottom-up visual input, the accuracy was higher in normally sighted individuals than in individuals with reversed congenital cataracts. When bottom-up input was removed, decoding of categories remained successful in both groups, but decoding of individual scenes was only possible in normally sighted control individuals. Connective field modeling results indicated a less precise alignment of feedforward and feedback processing during visual stimulation in individuals with reversed congenital cataracts. These findings suggest that early visual experience is crucial for the refinement of feedback activity which in turn is crucial for well-tuned feedforward processing.

Population receptive field (pRF) mapping is widely used to characterize retinotopic organization based on functional magnetic resonance imaging (fMRI) data. Despite its broad adoption, the factors governing intra- and inter-subject variability in pRF estimates remain incompletely understood, limiting the ability to evaluate and optimize visual stimulation paradigms prior to data collection. Here, we investigate whether large-scale simulations can reproduce in vivo run-to-run variability patterns observed in pRF mapping and provide mechanistic insight into their origins. With GEMSim-pRF, our newly proposed computational framework for large-scale simulation and estimation of pRF responses, we generated millions of synthetic fMRI time courses across a wide range of receptive field parameters and noise conditions. We analyzed the variability of pRF estimation results derived from simulations and compared them with in vivo data from the publicly available NYU Retinotopy Dataset. Here we show that our simulation results matched the characteristic eccentricity-dependent variability observed in empirical pRF estimates. These findings show that key variability patterns observed in empirical pRF mapping can be successfully reproduced in large-scale simulations, establishing simulation-based analysis as a practical approach for understanding, predicting, evaluating and ultimately improving the behaviour of retinotopic mapping paradigms before empirical data collection.

Fixation is often treated as a period of stable visual processing. Yet, fixation is often punctuated by frequent microsaccades that occur during tasks involving complex foveal stimuli. These small eye movements are preceded by changes in visual sensitivity, both at the upcoming movement goal and at the currently fixated location. However, previous work has largely focused on isolated stimuli, leaving unclear whether pre-microsaccadic modulations reflect changes in sensitivity alone or also alter the spatial interactions that govern object recognition. Visual crowding provides a direct test of this question because it depends on the integration and segregation of nearby features and constrains recognition even within the foveola. Using high-precision Dual Purkinje Image eye tracking with retinally contingent stimulus delivery, we measured acuity and crowding thresholds at the preferred locus of fixation (PLF), the starting point of the impending gaze shift, while observers either maintained fixation or prepared to execute a microsaccade to a cued location. Unflanked acuity at the PLF remained stable across conditions. In contrast, crowding strength increased during the pre-microsaccadic interval, indicating an expansion of the foveal crowding zone. These results show that microsaccade preparation alters spatial integration at the starting point of the movement, increasing crowding even when sensitivity to isolated stimuli remains unchanged. Thus, microsaccades reshape foveal vision not only by modulating visual discrimination at the movement goal, but also by changing how nearby features are integrated and segregated before the eyes move. Significance StatementVision is often assumed to be most stable when gaze is fixed. Yet the eyes are never truly still, and the brain continually prepares small movements that shape perception before they occur. This study shows that such preparation changes how visual information is organized at the very center of gaze. Upcoming eye movements do not simply alter sensitivity to isolated objects; instead, they change how nearby features are integrated. These findings reveal that fine spatial vision and object recognition depends not only on what falls on the retina or on subsequent cortical processing but also on what the eyes are preparing to do next.

Prestimulus pupil size is associated with near-threshold detection performance in both the visual and auditory domain, a relationship that is commonly attributed to arousal. However, given that larger pupils also let more light into the eye, in the visual domain this relationship is likely also driven by optical effects. To better understand this, we investigated how pupil size, skin conductance, and electroencephalographic (EEG) measures relate to detection performance in both a visual and an auditory task. We found that larger pupils were associated with higher sensitivity in the visual condition but lower sensitivity in the auditory condition. Skin conductance was negatively related to visual sensitivity, but unrelated to auditory sensitivity. EEG power measures were not related to sensitivity in either condition, though pupil size was positively correlated with alpha and beta power. Together, these results suggest that in visual detection the relationship between pupil size and performance is driven by both arousal and optics, whereas in auditory detection the relationship is driven solely by arousal. More broadly, our findings highlight the pupil as an active and functional component of the visual system.

The critical flicker fusion (CFF) threshold is a psychophysical measure used to quantify the temporal resolution of the visual system and is known to vary across individuals. However, it is unclear if this measure is stimulus-specific, or if it may represent a more fundamental processing rate for visual perception in general. Here, we assess if individual variation in CFF is predictive of two features of visual processing that are dependent on temporal perception: the attentional blink and global motion sensitivity. In a non-clinical sample of 84 individuals, flicker fusion thresholds were predictive of the magnitude of the attentional blink. In contrast, we found no link between flicker fusion and global motion sensitivity in a sample of 79 individuals. Our results suggest that CFF reflects a visual processing rate that impacts other, more complex perceptual tasks.

PurposePupil diameter in daily life depends on both the light reaching the eye and the observers age, but established prediction formulas require laboratory quantities that are rarely measured in natural environments. We developed a compact age-corrected model that predicts pupil diameter from melanopic equivalent daylight illuminance (mEDI). MethodsWe used an existing field dataset in which binocular pupil diameter and near-corneal spectral irradiance were recorded while 83 adults aged 18-87 years moved through indoor and outdoor environments. The analysis included 10,082 valid paired observations. We fitted a bounded sigmoid relating pupil diameter to mEDI and age, with each participant given equal influence, and assessed prediction in participants excluded from model fitting. Performance was compared with simpler models, a flexible generalised additive model (GAM), and Watson-Yellott predictions based on assumed field geometry. ResultsPupil diameter decreased smoothly as mEDI increased. Age primarily reduced the difference between pupils in dim and bright conditions, by 0.768 mm per decade, while the predicted bright-light diameter changed little with age. In held-out participants, the bounded model had a participant-balanced root mean squared error (RMSE) of 0.630 mm and mean absolute error of 0.537 mm. The GAM had a slightly lower point-estimate RMSE of 0.610 mm, but the difference was small and uncertain. The bounded model outperformed the tested log-linear, reduced, age-only, and Watson-Yellott alternatives. ConclusionAge and mEDI are sufficient to provide useful population-average pupil predictions across the observed adult age and real-world light range. The model is transparent, physiologically bounded, and nearly as accurate as a flexible GAM, but predictions approaching darkness remain uncertain because valid mEDI measurements were not available in that range. Key pointsO_LIA compact equation predicts population-average pupil diameter from age and mEDI alone. C_LIO_LIAge mainly compresses the pupils response range by reducing pupil diameter under dimmer conditions. C_LIO_LIPrediction error in unseen participants was close to that of a flexible GAM, without requiring a fitted smooth object. C_LIO_LIThe model is intended for the observed adult age and field-light range, not for extrapolation into darkness. C_LI

The temporal resolution of vision is seriously limited. However, the response to a very brief flash, called the impulse response, is already quite sluggish at the earliest stages of vision, potentially obscuring the true temporal resolution of the rest of the visual system. Faster monitors that produce briefer flashes are subject to diminishing returns because, by definition, they cannot elicit responses that are any briefer than the impulse response. Here, taking inspiration from previous attempts, we develop a novel technique for presenting flashes that elicit 'briefer-than-brief' visual responses. Using a simple deconvolution technique, we reverse-engineer the visual response and estimate the form that the stimulus should take to elicit the response that a faster visual system would produce to a normal flash. Using psychophysics on human observers, we demonstrate that these 'briefer-than-brief' (BTB) flashes partially bypass the temporal limits presumably imposed by the early visual system using two paradigms: one that requires temporal segregation and one that requires temporal integration of sequential flashes. With BTB flashes, human observers successfully isolated two successive flashes at shorter intervals than with conventional flashes, improving temporal resolution by around 16%. We found that BTB stimuli not only improved temporal resolution, but also induced poorer performance on tasks requiring temporal integration, suggesting that the visual responses elicited by BTB flashes overlap less in time due to their briefer duration. In sum, our findings suggest that, using reverse-engineered stimuli, we can alleviate a temporal bottleneck that probably originates from the earliest stages of vision. In doing so, we allow higher visual areas to operate at a higher temporal resolution than previously thought possible.

Purpose: To evaluate the feasibility, validity, and scalability of PLFest, an open-source, Unity-based, cross-platform application designed for standardized, multi-site visual and cognitive assessment and training. Methods: Two hundred sixty participants (mean age = 23 years) were recruited across four university sites in the United States. Participants completed a battery of five visual assessments administered through PLFest, including visual acuity, contrast sensitivity, spatial frequency cutoff, contrast sensitivity at spatial-frequency cutoff, and visual search. Five cognitive assessments measuring visuospatial working memory, verbal working memory, fluid reasoning, inhibitory control, and selective attention were also administered. Descriptive statistics and performance distributions were examined and compared with normative data. Results: Visual acuity and contrast sensitivity measures closely matched previously reported normative values obtained using established clinical and psychophysical methods. Spatial frequency cutoff and visual search tasks produced stable threshold estimates while showing substantial inter-individual variability. Performance across all cognitive assessments was consistent with published validation studies of the corresponding tasks. Across the full battery, adaptive procedures demonstrated reliable convergence and generated well-distributed performance measures without evidence of substantial floor or ceiling effects. Importantly, these findings were observed across four geographically distributed testing sites using standardized consumer-grade tablet hardware. Conclusions: PLFest provides reliable and scalable assessment of visual and cognitive function using portable consumer devices. The platform supports standardized data collection across distributed research settings while maintaining performance characteristics consistent with established laboratory and clinical benchmarks. These findings support the use of PLFest as a reliable framework for large-scale studies of vision and cognition. Translational Relevance: By reducing dependence on specialized laboratory infrastructure and trained personnel, PLFest may facilitate broader access to visual and cognitive assessment, enabling large-scale research, screening, and future rehabilitation applications.

Perception is typically biased towards prior expectations. In some cases, however, it seems repelled away from expectations, such that percepts appear less like what is expected. Even more intriguingly, separate studies have recently reported that predictions derived from gravity may shape perception in opposing ways. Specifically, gravity causes unsupported objects to accelerate downwards, leading to two predictions; that objects will move downwards (location prior) and at an increasing speed (acceleration prior). There is evidence that perceptual judgements are attracted towards location priors yet repelled from acceleration ones. Here we examine these effects in the same paradigm to determine whether they result from different types of stimuli and judgement, or more interestingly, might result from opposite influences of common predictive mechanisms influencing perception. We first replicate previous reports of a systematic bias to report upward moving objects as more accelerating than downward moving objects: effectively a repulsion from acceleration priors. We then show that the effect applies both at the level of retinal space and due to contextual cues concerning gravitational direction. Finally, we find that participants errors in a location reproduction task are similarly consistent with a repulsion from acceleration priors and, simultaneously, with an attraction towards location priors. We conclude by considering the ways in which these concurrent attractive and repulsive biases may reflect mechanisms optimising fast, accurate, and informative experiences in our ever-changing sensory world, therefore optimising the interface between perception and learning. Public Significance StatementIn a series of behavioural experiments, we show that expectations about how objects move due to gravity concurrently attract perception towards the prediction that objects move downwards, and repel perception away from the prediction that they do so at an increasing rate. These opposing influences inform current theories of perceptual processing, which explain how expectations may generate percepts that are fast, veridical, and informative.

Generalization of visual perceptual learning (VPL) to unseen conditions varies across tasks. Previous work suggests that training curriculum may be integral to generalization, yet a theoretical explanation is lacking. We propose an explanatory theory of visual learning generalization and curriculum effects by leveraging an artificial neural network (ANN) model of VPL in comparison with humans. We found that easy-to-hard sequential training improved generalization in both humans and ANNs. However, when easy and hard conditions were interleaved, humans and ANNs showed different behaviours: while ANNs performed worse than with sequential training, humans maintained good performance but with large inter-individual variability. Investigating ANN models trained with different curricula, we demonstrated that models relying on low-dimensional neural populations showed superior generalization. This readout subspace dimensionality was directly determined by curriculum: learners who learned from easy tasks early formed lower-dimensional subspaces and generalized better. Our theory provides a mechanistic framework linking curriculum design to VPL generalization through neural population dimensionality. Author SummaryLearning new skills is fundamental to humans and animals. However, a key challenge in learning is generalization: applying learned skills to new situations not encountered during training. While it is well known that training curriculum affects how well we generalize, the underlying mechanisms remain poorly understood. What makes certain training curricula more effective than others? Why do some learners generalize better than others even with similar training? We developed a computational theory to explain how curriculum design influences generalization in visual learning. Using a combination of human behavioural experiments and artificial neural network modeling, we posit that easy-to-hard training sequences lead learners to focus on fewer, more essential visual features. This creates low-dimensional neural representations that are robust and generalize well to new conditions. Importantly, we found this benefit even extends to learners who spontaneously adopt easy-to-hard strategies on their own, without explicit curriculum design. Our theory provides a mechanistic link between curriculum structure, neural population dynamics, and generalization performance. Moreover, these findings offer practical guidelines for designing effective training programs in education and rehabilitation.