This thesis investigates perceptual alignment, the degree to which deep learning models process visual information in ways similar to human perception. While modern neural networks achieve exceptional performance in tasks such as segmentation, classification, and multimodal reasoning, their internal computations often diverge significantly from those of the human visual system. The goal of the thesis is to understand when, why, and how artificial vision systems behave like (or unlike) humans, and to evaluate whether biologically inspired mechanisms can bridge this gap. The work is structured around three main objectives: 1. Integrating bio-inspired mechanisms into deep architectures: The thesis implements divisive normalization (DN), a canonical computation observed in the early visual cortex, within U-Net segmentation models. DN acts as a gain-control mechanism, modulating responses based on local contrast. Experiments reveal that DN improves robustness in adverse conditions such as fog, low light, and reduced contrast, with minimal computational overhead. However, these improvements in robustness do not translate into more human-like perceptual behavior. 2. Developing psychophysically inspired evaluation methods: To measure perceptual alignment, the thesis proposes systematic frameworks inspired by human psychophysics. These include: - A Decalogue of low-level visual tests (contrast sensitivity, masking, adaptation, etc.). - An adaptation of MacAdam ellipses to evaluate chromatic discrimination in neural networks. - Contrast sensitivity function (CSF) analyses for multimodal models. - A layer-wise study of abstraction in CLIP, showing how linguistic supervision shifts representations from texture-based (more human-like early on) to semantic/global-shape-based (less aligned with human low-level perception). These tools show that perceptual alignment varies widely across architectures, layers, training regimes, and data distributions. 3. Identifying the factors driving (mis)alignment: A broad comparison across CNNs, Vision Transformers, CLIP, and multimodal language models shows that alignment is multifactorial. Crucially: - The statistics of the training data (e.g., chromatic richness) strongly influence perceptual properties. - Linguistic supervision pushes representations toward high-level abstractions, reducing sensitivity to low-level cues. - Increasing model scale, optimization, or regularization often reduces human alignment, producing an inverted-U relationship between accuracy and perceptual similarity. Overall, the thesis demonstrates that biological plausibility does not guarantee human-like behavior, and high task performance can actually worsen alignment. Perceptual alignment emerges not from architecture alone but from the interplay of data, objectives, and supervision. The work concludes that closing the gap between artificial and human vision requires new training objectives, richer evaluation frameworks, and human-centric design principles informed by psychophysics and neuroscience.
