A Turing Test for Artificial Nets devoted to model Human Vision
summary
The gist
In this work we argue that, despite recent claims about successful modeling of the visual brain using deep nets, the problem is far from being solved…
In short
The paper proposes a low-level Turing test using ten specific physiological and psychophysical facts (the Decalogue) to evaluate if artificial deep learning models accurately mimic the human visual brain. Testing three different models showed that one, tuned via Maximum Differentiation, performed best in nonlinear responses to complex patterns, suggesting current deep learning methods are insufficient for low-level vision modeling.
Key concepts
- Decalogue
- A set of ten specific low-level physiological and psychophysical facts describing the adaptive information bottleneck of the retina-V1 pathway. These facts cover spectral sensitivities, color response saturation, achromatic contrast sensitivity, and various nonlinear responses to texture and context effects.
- Low-level Turing Test
- An evaluation framework where an artificial model is tested by its ability to correctly respond to a set of fundamental human visual facts (the Decalogue). If the model's behavior matches the human response to these basic physical principles, it is considered a better mimic of the low-level visual brain.
- Nonlinear Contrast Response
- The way humans perceive contrast changes nonlinearly depending on factors like background energy or feature similarity. This concept describes how our sensitivity to contrast isn't a simple linear measure; it involves saturation, masking, and context-dependent changes in perception.
Terminology used across episodes
This episode discusses
- A Turing Test for Artificial Nets devoted to model Human Vision · Paper Radio
- MindSet: Vision. A toolbox for testing DNNs on key psychological experiments
- Do computer vision foundation models learn the low-level characteristics of the human visual system?
- Evaluating quality metrics through the lenses of psychophysical measurements of low-level vision
- Brain-Like Object Recognition with High-Performing Shallow Recurrent ANNs
- End-to-end Optimized Image Compression
- Parametric PerceptNet: A bio-inspired deep-net trained for Image Quality Assessment
The paper
A Turing Test for Artificial Nets devoted to model Human Vision · Read on arXiv
Universitat de Valenciáa
In our invited talk at the AI Evaluation Workshop of the University of Bristol back in June 2022 we argued that, despite claims about successful modeling of the visual brain using ANNs, the problem is far from being solved (even for low-level vision). Open issues include: where should we read from ANNs to reproduce human behavior?, this ad-hoc read-out is part of the brain model or not?, should we use artificial psychophysics or artificial physiology?, artificial experiments should literally match the experiments in humans?. There is a clear need of rigorous procedures for experimental tests for ANNs models of the visual brain, and more generally, to understand ANNs devoted to generic vision tasks. Following our experience in using low-level facts from Visual Neuroscience in Image Processing, we presented the idea of developing a low-level dataset compiling the basic spatio-temporal and chromatic facts that are known to happen in the retina-V1 pathway, and they are not currently available in existing databases such as BrainScore. In our results we checked the behavior of three recently proposed models with similar architecture: (1) A parametric model tuned via Maximum Differentiation [Malo & Simoncelli SPIE 15, Martinez et al. PLOS 18, Martinez et al. Front. Neurosci. 19], (2) A non-parametric model, the PerceptNet, tuned to maximize the correlation with human opinion on subjective distortions [Hepburn et al. IEEE ICIP 20], and (3) A model with the same encoder as PerceptNet, but tuned for segmentation (published later as Hernandez-Camara et al. Patt.Recogn.Lett. 23, Hernandez-Camara et al. Neurocomp. 25). Results on 10 compelling psycho/physio visual facts show that the first model is the one with closer behavior to the humans in terms of receptive fields, but more interestingly, on the nonlinear behavior for spatio-chromatic patterns of a range of luminances and contrasts.
DOI: 10.3389/frai.2025.1665874
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "A Turing Test for Artificial Nets devoted to model Human Vision".
Marcus: Here's the gist of the paper. In this work we argue that, despite recent claims about successful modeling of the visual brain using deep nets, the problem is far from being solved, particularly for low-level vision The key ideas of this work are basically two: • The use of AI techniques (e.g.
Ines: deep learning) to understand the visual brain may not be as easy as people thought back in 2022, and even now • Our specific proposal here is a low-level Turing test based on 10-points low-level physiological and psychophysical facts (our Decalogue) to check if certain artificial model behaves as the (low-level) human visual brain.
Marcus: The Proposed Evaluation Framework The authors propose a low-level Turing test based on 10 compelling psycho/physio visual facts, which they call their Decalogue.
Ines: This set of facts is compiled from low-level facts from Vision Science in image processing that describe the adaptive information bottleneck of the retina-V1 pathway.
Title and authors: Ines: So we’re looking at this paper now called "A Turing Test for Artificial Nets devoted to model Human Vision," and the title immediately tells us they’re proposing a test—a Turing Test—specifically to check if deep nets actually match how the visual brain works at the lowest level.
Marcus: Exactly. It's not about whether an AI can classify a picture or do a high-level task, it’s about those fundamental facts, the low-level physiological and psychophysical facts that describe how your retina and V1 pathway interact.
Yuki: From my side, I wonder what this test really means for us as a population geneticists; can these low-level checks actually support claims about how complex visual systems evolve across species?
Ines: That’s a fair question. The paper is focused on establishing an evaluation framework that uses these specific facts to see if the models are behaving like humans in terms of basic visual processing, not just achieving high accuracy on some abstract label.
Marcus: Right, so the core idea is moving away from just looking at final performance metrics and demanding that the AI exhibits specific, measurable low-level behaviors we observe in human vision.
The paper's summary: Ines: The paper explains that despite a lot of talk about deep nets modeling the visual brain, they argue the problem isn't solved yet, especially when we look at low-level vision. They propose their own specific test set, which they call their Decalogue.
Marcus: That decalog is essentially a set of ten compelling psycho-physiological facts compiled from image processing vision science that describe how information is adaptively captured and discarded by the retina-V1 pathway.
Yuki: Ten specific facts sound very detailed; I’m curious if this level of detail can actually be generalized to support broader claims about visual processing in different contexts, not just static images.
Ines: The authors generate digital images illustrating these ten facts, and they feed them into any artificial image-computable model to see how the model reacts.
Marcus: They check three different recent models with similar deep architectures—a parametric one tuned by Maximum Differentiation, a non-parametric one called PerceptNet, and another optimized for segmentation—to see which one aligns best with human behavior on these ten facts.
The paper's improvements: Ines: The results show that the parametric model, the first one they tested, actually had closer behavior to humans when it came to nonlinear response when facing complex spatio-chromatic patterns.
Marcus: That’s important because the analysis shows that while that parametric model aligned better in linear parts, like sensitivity and Gabor receptive fields, other models had issues; for example, the BioMultiLayer network showed human-like spectral sensitivity at the front end but it degraded throughout the network.
Yuki: So it seems even with a good architecture like PerceptNet or bio-inspired designs, if they aren't tuned for these specific low-level nonlinearities, they still fall short of what humans do when processing complex visual input.
Ines: The authors point out that the qualitative evaluation suggests we need architectural changes—like using Divisive Normalization to impose interaction—or changing the training goals to get more plausible receptive fields in the AI.
Marcus: They also suggest that having robust priors related to natural image statistics can benefit any network trying to model vision, which means they see this low-level Turing test as a way for researchers to question the tasks or restrictions they use when training these models.
Conclusion: Ines: To wrap up, the paper with this title, "A Turing Test for Artificial Nets devoted to model Human Vision," demonstrates that even with just a qualitative check using their Decalogue, a significant ranking is possible—the parametric model showed bigger alignment in the linear parts of the facts.
Marcus: They conclude that any researcher devoted to vision should see how their models respond to this test because it helps question the training data and tasks used.
Yuki: For me, as someone looking at population genetics, these findings imply that if we want AI models to actually function on biological visual tasks, we need to incorporate those specific adaptive bottlenecks right at the very start of their design process.
Ines: It’s a call for iterative optimization where you alternate between different goals and see how they interact, suggesting that these low-level facts provide essential priors for any network intended for vision.
Marcus: So, to summarize, this work uses the Decalogue to show that simply building a deep net isn't enough; we need targeted evaluation against human visual physiology to ensure the AI is actually capturing the right biological mechanisms.
More episodes
- 2610.33881-Sensorimotor behavior from an untrained brain-and-nerve-cord connectome
- 2610.40344-Morphological Features of Murine Pulmonary Arterial Networks for Diagnosing Vascular Disease
- Daily Summary for 2026-11-19
- 2609.10193-Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- 2610.38236-Generative discovery of chemical reaction networks
- 2507.11176-An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2504.03590-Towards a Unified Framework for Determining Conformational Ensembles of Disordered Proteins
- 2603.15217-A multiscale discrete-to-continuum framework for structured population models