Texture Representations in Deep Neural Models and Humans

Visual perception is often studied through the lens of object recognition. However, a large fraction of visual experience concerns image properties that are not naturally described in terms of discrete objects or categories. Visual textures provide a useful method to investigate this complementary aspect of vision, as they can be described through image statistics while spanning a wide range of perceptual and semantic complexity. In this thesis, I investigate how textures are represented in deep neural networks and humans, combining texture synthesis, representation analysis, computational modeling, and psychophysics. I first explored whether the naturalistic texture representations introduced by Gatys et al. (2015) could be combined with the latent space of a Variational Autoencoder to construct a controllable generative model of naturalistic textures. Although the complete model did not achieve the intended performance, the experiments isolated the texture representation proposed by Gatys as the most robust component of our framework and motivated a direct investigation of its representational properties. I then studied texture representations coded as Gram matrices across a diverse set of convolutional neural networks trained for object recognition. Texture information increased with network depth, but the resulting representations recovered only part of the categorical structure of natural textures defined by human judgment. Crucially, their quality was not correlated with measures of neural or behavioral alignment in object recognition. This dissociation indicates that being a good model of the ventral visual stream in an object recognition setting does not necessarily imply representing textures in a human-like way. Finally, I extended this analysis beyond convolutional architectures by comparing one CNN with three Vision Transformers across textures of progressively increasing complexity and with human performance in a psychophysical discrimination task. Vision Transformers formed substantially more consistent representations across texture types than the CNN, and these representational relationships were significantly better aligned with human perception. Taken together, these results show that texture perception provides a complementary benchmark for computational models of vision and exposes limitations within the dominant object recognition paradigm. More broadly, I suggest that understanding biological visual representations requires models and evaluation criteria that account not only for object identity and semantic content, but also for the statistical and contextual structure through which visual appearance is organized.