Understanding the Visual Cortical Pathways From a Computational Perspective

A longstanding two visual cortical streams hypothesis suggests that object properties (what) and spatial properties (where) are processed separately in different cortical visual pathways. This hypothesis has been supported by much evidence in neuroscience studies. However, there are some conflicting findings regarding the two-stream hypothesis in the last few decades that demand explanation. For example, shape activities (fMRI) and selectivities (physiology) have been found in both ventral and dorsal pathways; likewise, spatial activations (fMRI) and spatial selectivities (physiology) have also been found in both ventral and dorsal pathways. Previous studies also suggest that there are multiple segregated sub-pathways within a main visual pathway. In addition, processing different visual attributes of multiple objects in a scene using segregated pathways raises questions about how the brain combines different visual attributes of each object together to form an integrated view of each object (the binding problem). We introduce the history of the two-stream hypothesis and discuss experimental evidence that challenge our current understanding of the two-stream hypothesis. Based on our computational modeling studies, we propose some explanations about why the brain has segregated pathways, when the brain should develop a segregated pathway for a visual attribute, and how the brain constrains the binding problem. Our computational modeling studies about the two-stream hypothesis might help us design better computer vision models. Finally, we summarize our conclusions and discuss the potential future directions that can be explored in the future.