Computational neuropsychology 2.0: A deep learning account of the pattern of deficits in visual recognition after brain damage

Localized brain damage can result in relatively specific problems in
mental faculties, including problems with visual recognition. It is still a
matter of debate why certain deficits and combinations of deficits
occur. Recent evolutions in human neuroscience, including our own
work, have revealed the complexity of the system that supports
visual recognition, and models such as deep neural networks (DNNs)
have offered us a computational understanding of many properties of
this system. This empirical knowledge and computational insights
have high potential for a better understanding of neuropsychological
deficits, and, in the other direction, neuropsychological cases offer a
unique approach to validate the available computational models. In
this project we will apply this new approach, here referred to as
computational neuropsychology 2.0. We will develop and validate a
new test battery of recognition tasks including multiple categories
and a manipulation of recognition difficulty. We will show how
performance associations between object categories in healthy
individuals and neural overlap in the healthy brain is related to
representational overlap in DNNs. Crucially, we will test a large
sample of stroke patients, and show to what extent the pattern of
deficits across object categories and across difficulty levels can be
explained by the information processing in DNNs, allowing us to
'mimic' the full spectrum of human recognition deficits in artificial
intelligence models.