Biology subjects

Maruya, A.

Publications and source records attributed to Maruya, A..

6 recordsLinked to original sources

bioRxiv2026

Foveation as self-supervision signal: Blur masks beat blank masks for learning robust visual representations

The primate retina is strikingly non-uniform, with receptor density falling off sharply from fovea to periphery. Every eye-movement brings a previously low-acuity peripheral area to high-acuity foveal processing. We hypothesize that this process provides a natural self-supervision signal, enabling representation learning by predicting fine details in the periphery. We implement this in a SimMIM-style vision transformer, replacing its blank masks (ViT-Blank) with a uniform blur mask (ViT-Blur) and further with retinal masks in which a foveal patch is randomly selected, and all other patches are progressively blurred based on eccentricity from the foveal patch (ViT-Retina). ViTs were pre-trained to reconstruct the full-resolution image from the corrupted input and then fine-tuned on clean images for classification. ViT-Blur and ViT-Retina outperform ViT-Blank on the pre-training reconstruction task while controlling for retained information from the input across corruption types. Critically, they also outperform ViT-Blank on downstream classification, showing that they learn representations that generalize better. The ViT-Blur and ViT-Retina are also more robust in impoverishment conditions such as higher mask ratios, fewer pre-training epochs, smaller training sets, and test time image corruptions. Using additive band-limited noise, we discovered that ViT-Blank relies on higher spatial frequencies compared to ViT-Blur and ViT-Retina, and we show this to be the likely cause of poorer generalization: the centroid of a model's spatial frequency tuning is negatively correlated with its classification accuracy ($r = -0.73$). We further show that a closed-form linear-ridge encoder derived from the same reconstruction objective reproduces this tuning ($r = 0.88$). Overall, our experiments show the promise of a biologically-inspired self-supervision objective for learning robust visual representations.

neuroscience↗

bioRxiv2025

Generative Phenomenology of form perception: Perceptograms and cortical models for amblyopic form distortions

What we see is generated by our cortical neurons processing sensory input. Amblyopia provides a perfect case for studying the neural generation of percepts because abnormal neural development of cortex leads to many amblyopes seeing more complex phantom forms than the viewed stimulus through their amblyopic eye (AE) but not through the fellow eye (FE). Using computer guided dichoptic displays, we showed that 92.6% of phantom percepts generated by single gratings shown to AE were matched exactly by plaids (sums of two gratings) shown to FE. A formal equation between the cortical signals generated by the gratings seen through AE and the signals generated by their matched patterns seen through FE, provided the derivation of AE cortical receptive fields as linear transforms of FE receptive fields. The transformed AE cortical model accurately generated the phantoms, was validated via reverse-correlation, and explained deficits in form detection and spatial localization. This study elucidates the close link between percepts and neuronal receptive fields. TEASERChanges in cortical neurons receptive fields can lead to amblyopes seeing phantom features that are not present in viewed images.

neuroscience↗

bioRxiv2024

Anisotropy of object non-rigidity: High-level perceptual consequences of cortical anisotropy

We demonstrate an unexpected anisotropy in perceived object non-rigidity, a higher-level perceptual phenomenon, and explain it by the population distribution of low-level neuronal properties in primary visual cortex. We measured the visual interpretation of two rigidly connected rotating circular rings. In videos where observers predominantly perceived rigidly-connected horizontally rotating rings, they predominantly perceived non-rigid independently wobbling rings if the video was rotated by 90{degrees}. Additionally, vertically rotating rings appeared narrower and longer than horizontally rotating counterparts. We decoded these perceived shape changes from V1 outputs incorporating documented cortical anisotropies in orientation selectivity: more cells and narrower tuning for the horizontal orientation than for vertical. Even when shapes were matched, the non-rigidity anisotropy persisted, suggesting uneven distributions of motion-direction mechanisms. When cortical anisotropies were incorporated into optic flow computations, the kinematic gradients (Divergence, Curl, Deformation) for vertical rotations aligned more with derived gradients for physical non-rigidity, while those for horizontal rotations aligned closer to rigidity. Our results reveal how high-level non-rigidity percepts can be shaped by hardwired cortical anisotropies. Cortical anisotropies are claimed to promote efficient encoding of statistical properties of natural images, but their surprising contribution to failures of shape constancy and object rigidity raise questions about their evolutionary function.

neuroscience↗

bioRxiv2023

Object Rigidity: Competition and cooperation between motion-energy and feature- tracking mechanisms and shape-based priors

Why do moving objects appear rigid when projected retinal images are deformed non-rigidly? We used rotating rigid objects that can appear rigid or non-rigid to test whether shape features contribute to rigidity perception. When two circular rings were rigidly linked at an angle and jointly rotated at moderate speeds, observers reported that the rings wobbled and were not linked rigidly but rigid rotation was reported at slow speeds. When gaps, paint or vertices were added, the rings appeared rigidly rotating even at moderate speeds. At high speeds, all configurations appeared non-rigid. Salient features thus contribute to rigidity at slow and moderate speeds, but not at high speeds. Simulated responses of arrays of motion-energy cells showed that motion flow vectors are predominantly orthogonal to the contours of the rings, not parallel to the rotation direction. A convolutional neural network trained to distinguish flow patterns for wobbling versus rotation, gave a high probability of wobbling for the motion-energy flows. However, the CNN gave high probabilities of rotation for motion flows generated by tracking features with arrays of MT pattern-motion cells and corner detectors. In addition, circular rings can appear to spin and roll despite the absence of any sensory evidence, and this illusion is prevented by vertices, gaps, and painted segments, showing the effects of rotational symmetry and shape. Combining CNN outputs that give greater weight to motion energy at fast speeds and to feature tracking at slow, with the shape-based priors for wobbling and rolling, explained rigid and nonrigid percepts across shapes and speeds (R2=0.95). The results demonstrate how cooperation and competition between different neuronal classes leads to specific states of visual perception and to transitions between the states.

neuroscience↗

bioRxiv2020

Mental geometry of perceiving 3D size in pictures

We show that the classical problem of 3D size perception in obliquely viewed pictures can be understood by comparing human performance to the optimal geometric solution. A photograph seen from the camera position, forms the same retinal projection as the physical 3D scene, but retinal projections of sizes and shapes are distorted in oblique viewing. For real scenes, we previously showed that size and shape inconstancy result despite observers using the correct geometric back-transform, because some retinal images evoke misestimates of viewing elevation. Now, we examine how observers estimate 3D sizes in oblique views of pictures of objects in different poses on the ground. Compared to the camera position estimates, sizes in oblique views were seriously underestimated for objects at fronto-parallel poses, but there was almost no change for objects perceived as pointing towards the viewer. The inverse of the function relating projected length to pose, camera elevation and viewing azimuth, gives the optimal correction factor for inferring correct 3D lengths if the elevation and azimuth are estimated accurately. Empirical correction functions had similar shapes to optimal, but lower amplitude. Measurements revealed that observers systematically underestimated viewing azimuth, similar to the fronto-parallel bias for object pose perception. A model that adds underestimation of viewing azimuth to the geometrical back-transform, provided good fits to estimated 3D lengths from oblique views. These results add to accumulating evidence that observers use internalized projective geometry to perceive sizes, shapes and poses in 3D scenes and their pictures.

neuroscience↗

bioRxiv2019

Mental geometry of 3D size and shape perception

Judging poses, sizes and shapes of objects accurately is necessary for organisms and machines to operate successfully in the world. Retinal images of 3D objects are mapped by the rules of projective geometry, and preserve the invariants of that geometry. Since Plato, it has been debated whether geometry is innate to the human brain, and Poincare and Einstein thought it worth examining whether formal geometry arises from experience with the world. We examine if humans have learned to exploit projective geometry to estimate sizes and shapes of objects in 3D scenes.\n\nNumerous studies have examined size invariance as a function of physical distance, which changes scale on the retina, but surprisingly, possible constancy or inconstancy of relative size seems not to have been investigated for object pose, which changes retinal image size differently along different axes. We show systematic underestimation of length for extents pointing towards or away from the observer, both for static objects and dynamically rotating objects. Observers do correct for projected shortening according to the optimal back-transform, obtained by inverting the projection function, but the correction is inadequate by a multiplicative factor. The clue is provided by the greater underestimation for longer objects, and the observation that they appear more slanted towards the observer. Adding a multiplicative factor for perceived slant in the back-transform model provides good fits to the corrections used by observers. We quantify the slant illusion with relative slant measurements, and use a dynamic demonstration to show the power of the slant illusion.\n\nIn biological and mechanical objects, distortions of shape are manifold, and changes in aspect ratio and relative limb sizes are functionally important. Our model shows that observers try to retain invariance of these aspects of shape to 3D rotation by correcting retinal image distortions due to perspective projection, but the corrections can fall short. We discuss how these results imply that humans have internalized particular aspects of projective geometry through evolution or learning, and how assuming that images are preserving the continuity, collinearity, and convergence invariances of projective geometry, supplements the Generic Viewpoint assumption, and simply explains other illusions, such as Ames Chair.

neuroscience↗