Abstract

The success of any kind of visual interface, from a smartphone screen, to a head-mounted display, relies on the fact that the eye is rather bad in seeing the artefacts of those technologies. For example, we do not see the matrix of red-green-blue subpixels or that video is in fact a sequence of frames presented in succession. Despite that we rely on the limitations of the visual system in any aspect of visual technologies, our level of understanding and computational modelling of those limitations is surprisingly limited.

In this talk I will propose how computation can help to integrate partial information we have on the limitations of the visual system and build a comprehensive model of those limitations. The challenge here is different than in most machine learning problems: the amount of data is very limited, expensive to collect, and impossible to predict with simple models. Some of the interesting problems are how to reuse existing psychophysical data, or how to automatically find and generate new measurement points, which would be the most beneficial for model discovery. The complexity of the models leads to computationally demanding training methods, which require HPC resources and efficient parallel implementations.