Abstract
Previous studies showed that processing input images using artificial neural networks with segregated artificial ventral and dorsal pathways had higher performance in object recognition and localization tasks (higher average accuracy, lower variance, and less required training time) than processing the input images with a single pathway. The better performance on object recognition and localization tasks with segregated pathways might explain why our brain separates visual processing into two cortical visual streams (what versus where or how). However, it is unclear whether it is always computationally advantageous to use segregated pathways. That is, will segregated pathways for different stimulus attributes of an input image always improve network performance by allowing segregated pathways to process different visual stimulus attributes separately? In the current study, we trained artificial convolutional neural networks using supervised learning to determine identity (shape), luminance, orientation, and location of multiple objects in a black background. There were three non-overlapping black and white objects (tops, pants, or shoes) at different locations in each image. The luminance of each object could be high or low and the orientation of each object could be up or down. We found that in order to achieve highest performance, it was very important to process identity and location information separately in different pathways. It was also advantageous to process orientation and luminance information separately, but the effects were smaller. Together, the artificial neural network findings suggest that different attributes of visual stimuli should be processed in segregated streams to achieve highest performance. The findings suggest that identity and location should be processed in segregated main streams, whereas orientation and luminance may be processed in segregated sub-streams. These findings are generally in agreement with current understanding of the brain’s organization and provide insight into how to design better artificial visual perception systems.