Fixed layer Convolutional Neural Network
Alexandros Kyrloglou · 2018
Fully trained convolutional neural networks are being used nowadays in various applications. But, can we get an understanding on how they actually work? A small step in that direction is taken in this paper. The filters of the first layers seem to be fairly straightforward and mostly are known and recognizable. If one makes a network with fixed weights on the filters, how is its performance compared with a fully trained one? And how is the training time influenced? This paper, answers this question by experimenting in a verification style CNN put in three different situations: first a fixed layer network, second a set layer network and third a fully trained CNN. It is shown that although similar results can be achieved with the three networks, a fully trained one still has superior performance; however, training time is increased as the number of fixed layers are increased.