The two visual streams hypothesis is a robust framework that has inspired many studies in the past three decades. One of the well-studied claims of this hypothesis is the idea that the dorsal visual pathway is involved in visually guided motor behavior, and it is operating with a short memory. Conversely, this hypothesis claims that the ventral visual pathway is involved in object classification, and it works using a long-term memory. In this study, we tested these claims by training identical recurrent neural networks to either perform viewpoint-invariant object classification (a task attributed to dorsal stream) or orientation classification (a task attributed to dorsal stream) and measured how much they rely on their memory in each task. Using a modified leaky-integrator echo state recurrent network, we found that object classification requires a longer memory compared to orientation classification. However, when we used long-short-term memory (LSTM) networks, we observed that object classification requires longer memory only in larger datasets. Accordingly, our results suggest that having a longer memory is advantageous in performing ventral stream’s tasks more than their dorsal counterparts, as was originally suggested by the two-streams hypothesis.
bioRxiv Subject Collection: Neuroscience