A Turing Test for artificial nets devoted to vision

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. Open issues include where should we read from in ANNs to check behavior? What should be the read-out? Is this ad-hoc read-out considered part of the brain model or not? In order to understand vision-ANNs, should we use artificial psychophysics or artificial physiology? Anyhow, should artificial tests literally match the experiments done with humans? These questions suggest a clear need for biologically sensible tests for deep models of the visual brain, and more generally, to understand ANNs devoted to generic vision tasks. Following our use of low-level facts from Vision Science in Image Processing, we present a low-level dataset compiling the basic spatio-chromatic properties that describe the adaptive bottleneck of the retina-V1 pathway and are not currently available in popular databases such as BrainScore. We propose its use for qualitative and quantitative model evaluation. As an illustration of the proposed methods, we check the behavior of three recent models with similar deep architectures: (1) A parametric model tuned via the psychophysical method of 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 humans on subjective image distortions [Hepburn et al. IEEE ICIP 20], and (3) A model with the same encoder as the PerceptNet, but tuned for image segmentation [Hernandez-Camara et al. Patt.Recogn.Lett. 23, Hernandez-Camara et al. Neurocomp. 25]. Results on the proposed 10 compelling psycho/physio visual properties show that the first (parametric) model is the one with behavior closest to humans.