Exploring the Learning of Face Selective Units in Deep Neural Networks

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
Because Arti cial Neural Networks are inspired by the brain, they can give neuroscientists a powerful tool to understand what happens in human brains. However, ANNs are often considered as black boxes. In this paper we have tried to understand how deep neural networks compare in units that select for faces when trained on various quantities of this category. The training of four networks is based on a weight factor (0, 0.2, 0.5, 0.8) which re ects the in uence of the occurrence of categories where one category consists of faces (a high occurrence) and the other of sieves, shredders and hovercrafts (low occurrence). This paper shows that the number of units that are signi cantly active for faces stays the same across these networks, but that the accuracy of category face becomes better with a higher weight factor. The accuracy for the notface category becomes worse with a higher weight factor. Furthermore, it is shown that in the training phase, the face category performs very well quickly with a high weight factor, but the performance for the notface category gets worse for a higher weight factor. Also, a relationship is shown between the number of face active units and the feature maps and that very early in the training phase a structure of the units that select signi cantly for faces becomes visible.
Faculteit der Sociale Wetenschappen