Modelling the symptoms of semantic dementia with deep neural networks

Lev Kiar Avberšek 

 Full article (pdf)  |   Written in Slovene  |  Published: 30. September 2020 |  Reads: 248

Abstract

Deep learning is a set of machine learning techniques. Hierarchical structure and many processing layers allow deep neural networks to make complex transformations of input data and learn representations on multiple levels of abstraction. The recent success of deep neural networks in the categorization of objects on images triggered a boom in their usage for modelling cognitive processes based on visual representations. Some characteristic errors that the deep neural networks make in categorization resemble symptoms of semantic dementia – a disorder in which learning of concepts is aggravated and rigid. In the present study, we have examined the potential of simulating symptoms of semantic dementia with deep neural networks. 45 healthy participants assessed typicality and pseudotypicallity of objects on images. These objects belonged to 5 superordinate categories, each of them containing 10 subordinate categories, used for training of the model and its evaluation. Three final versions of the model were evaluated; one with no dropout of neurons, one with 70% dropout and one with 90% dropout. The results showed that the accuracy of deep neural networks in categorization varies as a function of typicality and to as a function of the neuronal dropout. The pattern of errors that the model made with pseudotypical cases did not match the pattern typical for the patients with semantic dementia. However, the version with 70% dropout showed behavior that indicates some analogies with symptoms of semantic dementia. The study brought forth an exploratory introduction into studying the relationship between symptoms of semantic dementia and behavior of deep neural networks. Future research should focus on the examination of (behavioral and brain) representations of patients with semantic dementia.

Keywords

deep neural networks, deep learning, semantic dementia, categorization, computational model

Cite

Avberšek, L. K. (2020). Modeliranje simptomov semantične demence z globokimi nevronskimi mrežami [Modelling the symptoms of semantic dementia with deep neural networks]. Eksperimentator, 4, 58–66.