Expanding the high-level Bristol-MindSet with low-level Visual
Psychophysics for a better evaluation of Artificial Intelligence
Part 1: 1 September - 31 October 2024
Part 2: 18 June - 18 July 2025
Biography
Jesús Malo (1970) received the MSc and PhD degrees in Physics in 1995 and 1999 respectively, both from the Universitat de València for his work on Computational Visual Neuroscience with applications to Image Processing. He was the recipient of the Vistakon European Research Award in 1994 for his experimental work in Physiological Optics. In 2000 and 2001 he worked as Fulbright Postdoc at the Vision Group of the NASA Ames Research Center (with A.B. Watson), and as Visiting Researcher at the Lab of Computational Vision of the Center for Neural Science, New York University (with E.P. Simoncelli) in 2001 and 2013. Currently he is Full Professor of Vision Science at the Dept. of Optics of the Universitat de València, and he leads the research on Visual Neuroscience at the Image and Signal Processing Group of the Universitat de València. He has been Associate Editor of IEEE Trans. Im. Proc. (2009-2013) and now he is Academic Editor of PLOS ONE, Associate Editor of Frontiers in Neuroscience (speciality Perception Science), and regular member of the panel of NeurIPS and ICLR. He is member of the European Optical Society (EOS) and of the Asociación de Mujeres Investigadoras y Tecnólogas (AMIT). He is interested in models of low-level human vision, their relations with information theory and statistical learning, and their applications to image processing and vision science experimentation. Find out more at: http://isp.uv.es/excathedra.html .
Research Summary
The project combines the facts about human vision found by Experimental Psychology with the mathematics of Artificial Intelligence. This combination (which is the long-term agenda of both visiting Prof. Malo and the host Professors at UoB, Prof. Bowers and Prof. Santos) has two high-impact outcomes: (A) the development of more human perceiving machines, and (B) a more quantitative understanding of the behaviour of the human visual brain. Nowadays Machine Learning leads to artificial systems with human-like behaviour, but there is no consensus on the way to quantify the similarity with humans. The lack of quantitative comparisons designed to reveal flaws in the algorithms and architectures is one of the reasons why Artificial Vision remains unexplainable.
The goal of the project is complementing the high-level perceptual dataset of Prof. Bowers (the Bristol-MindSet) with a substantial range of low-level visual psychophysics related to successful image and video coding algorithms and to perceptual metrics to assess image and video quality. Low-level facts include luminance and colour nonlinearities, spatio-temporal bandwidth in opponent chromatic channels, and masking and adaptation effects that generate low-level visual illusions. Unification and harmonization of low-level visual psychophysics in a way that can be used by engineers and psychologists to evaluate algorithms is critical both for a quantitative understanding of the visual brain and to generate more-human machines.
The project includes two activities for post-grad students and phd students. On the one hand, (1) a 1 hour lecture for post-grad students on psychology in which Prof. Malo will introduce the extension of Prof. Bowers dataset: the Valencia-Bristol MindSet and the relation of low-level psycho-visual facts with low-level regularities in natural images and videos. On the other hand, (2) a hands-on activity (e.g. two sessions of 2.5 hours in a computer lab with a break) for PhD students of quantitative sciences (e.g. Comp. Sci., Physics, Maths, Psychol., Neurosci.) on the evaluation of Artif. Intell. algorithms using the Bristol-Valencia MindSet
You can contact Professor Malo Lopez's host Professor Jeff Bowers for further information.
