Exploring neural architectures for simultaneously recognizing multiple visual attributes

Abstract

Much experimental evidence in neuroscience has suggested a division of higher visual processing into a ventral pathway specialized for object recognition and a dorsal pathway specialized for spatial recognition. Previous computational studies have suggested that neural networks with two segregated pathways (branches) have better performance in visual recognition tasks than neural networks with a single pathway (branch). One previously proposed possibility is that two pathways increase the learning efficiency of a network by allowing separate networks to process information about different visual attributes separately. However, most of these previous studies were limited, considering recognition of only two visual attributes, identity and location, simultaneously with a restricted number of classes in each attribute. We investigate whether it is always advantageous to use two-pathway networks when recognizing other visual attributes as well as examine whether the advantage of using two-pathway networks would be different when there are a different number of classes in each attribute. We find that it is always advantageous to use segregated pathways to process different visual attributes separately, with this advantage increasing with a greater number of classes. Thus, using a computational approach, we demonstrate that it is computationally advantageous to have separate pathways if the amount of variations of a given visual attribute is high or that attribute needs to be finely discriminated. Hence, when the size of the computer vision model is limited, designing a segregated pathway (branch) for a given visual attribute should only be used when it is computationally advantageous to do so.