Significance of feedforward architectural differences between the ventral visual stream and DenseNet

Bryan Tripp

TL;DR: An approximation of primate ventral stream as a convolutional network performs poorly on object recognition, and multiple architectural features contribute to this.

Keywords: vision, primate, deep learning

Abstract: There are many differences between convolutional networks and the ventral visual streams of primates. For example, standard convolutional networks lack recurrent and lateral connections, cell dynamics, etc. However, their feedforward architectures are somewhat similar to the ventral stream, and warrant a more detailed comparison. A recent study found that the feedforward architecture of the visual cortex could be closely approximated as a convolutional network, but the resulting architecture differed from widely used deep networks in several ways. The same study also found, somewhat surprisingly, that training the ventral stream of this network for object recognition resulted in poor performance. This paper examines the performance of this network in more detail. In particular, I made a number of changes to the ventral-stream-based architecture, to make it more like a DenseNet, and tested performance at each step. I chose DenseNet because it has a high BrainScore, and because it has some cortex-like architectural features such as large in-degrees and long skip connections. Most of the changes (which made the cortex-like network more like DenseNet) improved performance. Further work is needed to better understand these results. One possibility is that details of the ventral-stream architecture may be ill-suited to feedforward computation, simple processing units, and/or backpropagation, which could suggest differences between the way high-performance deep networks and the brain approach core object recognition.