Pablo Hernández-Cámara

Pablo Hernández-Cámara, Jorge Vila-Tomás, Valero Laparra et al.

Subjective image quality measures based on deep neural networks are very related to models of visual neuroscience. This connection benefits engineering but, more interestingly, the freedom to optimize deep networks in different ways, make them an excellent tool to explore the principles behind visual perception (both human and artificial). Recently, a myriad of networks have been successfully optimized for many interesting visual tasks. Although these nets were not specifically designed to predict image quality or other psychophysics, they have shown surprising human-like behavior. The reasons for this remain unclear. In this work, we perform a thorough analysis of the perceptual properties of pre-trained nets (particularly their ability to predict image quality) by isolating different factors: the goal (the function), the data (learning environment), the architecture, and the readout: selected layer(s), fine-tuning of channel relevance, and use of statistical descriptors as opposed to plain readout of responses. Several conclusions can be drawn. All the models correlate better with human opinion than SSIM. More importantly, some of the nets are in pair of state-of-the-art with no extra refinement or perceptual information. Nets trained for supervised tasks such as classification correlate substantially better with humans than LPIPS (a net specifically tuned for image quality). Interestingly, self-supervised tasks such as jigsaw also perform better than LPIPS. Simpler architectures are better than very deep nets. In simpler nets, correlation with humans increases with depth as if deeper layers were closer to human judgement. This is not true in very deep nets. Consistently with reports on illusions and contrast sensitivity, small changes in the image environment does not make a big difference. Finally, the explored statistical descriptors and concatenations had no major impact.

Pablo Hernández-Cámara, Jorge Vila-Tomás, Paula Dauden-Oliver et al.

Autonomous driving is a challenging scenario for image segmentation due to the presence of uncontrolled environmental conditions and the eventually catastrophic consequences of failures. Previous work suggested that a biologically motivated computation, the so-called Divisive Normalization, could be useful to deal with image variability, but its effects have not been systematically studied over different data sources and environmental factors. Here we put segmentation U-nets augmented with Divisive Normalization to work far from training conditions to find where this adaptation is more critical. We categorize the scenes according to their radiance level and dynamic range (day/night), and according to their achromatic/chromatic contrasts. We also consider video game (synthetic) images to broaden the range of environments. We check the performance in the extreme percentiles of such categorization. Then, we push the limits further by artificially modifying the images in perceptually/environmentally relevant dimensions: luminance, contrasts and spectral radiance. Results show that neural networks with Divisive Normalization get better results in all the scenarios and their performance remains more stable with regard to the considered environmental factors and nature of the source. Finally, we explain the improvements in segmentation performance in two ways: (1) by quantifying the invariance of the responses that incorporate Divisive Normalization, and (2) by illustrating the adaptive nonlinearity of the different layers that depends on the local activity.

Pablo Hernández-Cámara, Valero Laparra, Jesús Malo

One of the key problems in computer vision is adaptation: models are too rigid to follow the variability of the inputs. The canonical computation that explains adaptation in sensory neuroscience is divisive normalization, and it has appealing effects on image manifolds. In this work we show that including divisive normalization in current deep networks makes them more invariant to non-informative changes in the images. In particular, we focus on U-Net architectures for image segmentation. Experiments show that the inclusion of divisive normalization in the U-Net architecture leads to better segmentation results with respect to conventional U-Net. The gain increases steadily when dealing with images acquired in bad weather conditions. In addition to the results on the Cityscapes and Foggy Cityscapes datasets, we explain these advantages through visualization of the responses: the equalization induced by the divisive normalization leads to more invariant features to local changes in contrast and illumination.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Alexandra Gomez-Villa, Pablo Hernández-Cámara, Muhammad Atif Butt et al.

Color serves as a fundamental dimension of human visual perception and a primary means of communicating about objects and scenes. As vision-language models (VLMs) become increasingly prevalent, understanding whether they name colors like humans is crucial for effective human-AI interaction. We present the first systematic evaluation of color naming capabilities across VLMs, replicating classic color naming methodologies using 957 color samples across five representative models. Our results show that while VLMs achieve high accuracy on prototypical colors from classical studies, performance drops significantly on expanded, non-prototypical color sets. We identify 21 common color terms that consistently emerge across all models, revealing two distinct approaches: constrained models using predominantly basic terms versus expansive models employing systematic lightness modifiers. Cross-linguistic analysis across nine languages demonstrates severe training imbalances favoring English and Chinese, with hue serving as the primary driver of color naming decisions. Finally, ablation studies reveal that language model architecture significantly influences color naming independent of visual processing capabilities.

Nuria Alabau-Bosque, Jorge Vila-Tomás, Paula Daudén-Oliver et al.

Playing games is inherently human, and a lot of games are created to challenge different human characteristics. However, these tasks are often left out when evaluating the human-like nature of artificial models. The objective of this work is proposing a new approach to evaluate artificial models via board games. To this effect, we test the color perception and color naming capabilities of CLIP by playing the board game Hues & Cues and assess its alignment with humans. Our experiments show that CLIP is generally well aligned with human observers, but our approach brings to light certain cultural biases and inconsistencies when dealing with different abstraction levels that are hard to identify with other testing strategies. Our findings indicate that assessing models with different tasks like board games can make certain deficiencies in the models stand out in ways that are difficult to test with the commonly used benchmarks.

Jorge Vila-Tomás, Pablo Hernández-Cámara, Valero Laparra et al.

Human vision models are at the core of image processing. For instance, classical approaches to the problem of image quality are based on models that include knowledge about human vision. However, nowadays, deep learning approaches have obtained competitive results by simply approaching this problem as regression of human decisions, and training an standard network on human-rated datasets. These approaches have the advantages of being easily adaptable to a particular problem and they fit very efficiently when data is available. However, mainly due to the excess of parameters, they have the problems of lack of interpretability, and over-fitting. Here we propose a vision model that combines the best of both worlds by using a parametric neural network architecture. We parameterize the layers to have bioplausible functionality, and provide a set of bioplausible parameters. We analyzed different versions of the model and compared it with the non-parametric version. The parametric models achieve a three orders of magnitude reduction in the number of parameters without suffering in regression performance. Furthermore, we show that the parametric models behave better during training and are easier to interpret as vision models. Interestingly, we find that, even initialized with bioplausible trained for regression using human rated datasets, which we call the feature-spreading problem. This suggests that the deep learning approach is inherently flawed, and emphasizes the need to evaluate and train models beyond regression.

Pablo Hernández-Cámara, Jesus Malo, Valero Laparra

A number of scientists suggested that human visual perception may emerge from image statistics, shaping efficient neural representations in early vision. In this work, a bio-inspired architecture that can accommodate several known facts in the retina-V1 cortex, the PerceptNet, has been end-to-end optimized for different tasks related to image reconstruction: autoencoding, denoising, deblurring, and sparsity regularization. Our results show that the encoder stage (V1-like layer) consistently exhibits the highest correlation with human perceptual judgments on image distortion despite not using perceptual information in the initialization or training. This alignment exhibits an optimum for moderate noise, blur and sparsity. These findings suggest that the visual system may be tuned to remove those particular levels of distortion with that level of sparsity and that biologically inspired models can learn perceptual metrics without human supervision.

Pablo Hernández-Cámara, Alexandra Gomez-Villa, Jose Manuel Jaén-Lorites et al.

Understanding how Multimodal Large Language Models (MLLMs) process low-level visual features is critical for evaluating their perceptual abilities and has not been systematically characterized. Inspired by human psychophysics, we introduce a behavioural method for estimating the Contrast Sensitivity Function (CSF) in MLLMs by treating them as end-to-end observers. Models are queried with structured prompts while viewing noise-based stimuli filtered at specific spatial frequencies. Psychometric functions are derived from the binary verbal responses, and contrast thresholds (and CSFs) are obtained without relying on internal activations or classifier-based proxies. Our results reveal that some models resemble human CSFs in shape or scale, but none capture both. We also find that CSF estimates are highly sensitive to prompt phrasing, indicating limited linguistic robustness. Finally, we show that CSFs predict model performance under frequency-filtered and adversarial conditions. These findings highlight systematic differences in frequency tuning across MLLMs and establish CSF estimation as a scalable diagnostic tool for multimodal perception.

Pablo Hernández-Cámara, Jose Manuel Jaén-Lorites, Jorge Vila-Tomás et al.

Vision Transformers (ViTs) achieve remarkable performance in image recognition tasks, yet their alignment with human perception remains largely unexplored. This study systematically analyzes how model size, dataset size, data augmentation and regularization impact ViT perceptual alignment with human judgments on the TID2013 dataset. Our findings confirm that larger models exhibit lower perceptual alignment, consistent with previous works. Increasing dataset diversity has a minimal impact, but exposing models to the same images more times reduces alignment. Stronger data augmentation and regularization further decrease alignment, especially in models exposed to repeated training cycles. These results highlight a trade-off between model complexity, training strategies, and alignment with human perception, raising important considerations for applications requiring human-like visual understanding.

Pablo Hernández-Cámara, Jose Manuel Jaén-Lorites, Jorge Vila-Tomás et al.

During the training of multi-modal models like CLIP, we observed an intriguing phenomenon: the correlation with low-level human image quality assessments peaks in the early epochs before gradually declining. This study investigates this observation and seeks to understand its causes through two key factors: shape-texture bias alignment and classification accuracy drop under noise. Our findings suggest that CLIP initially learn low-level visual features, enhancing its alignment with low-level human perception but also increasing its sensitivity to noise and its texture bias. As training progresses, the model shifts toward more abstract shape-based representations, improving noise robustness but reducing alignment with low-level human perception. These results suggest that these factors shared an underlying learning mechanism and provide new insights into optimizing the trade-off between perceptual alignment and robustness in vision-language models.