[1910.11105] Adaptive and Iteratively Improving Recurrent Lateral Connections
This view is an alternative view to that of using recursion for emulating deeper networks with shared weights (in our network the weights change but the same input u is reintroduced), to the views that emphasize top-down and bottomup integration (so far, we apply our method to the top layers and we focus on lateral connections), to explicit hypothesis verification models, and to attention as it is currently understood (our design is very different from attention blocks).
Abstract: The current leading computer vision models are typically feed forward neural
models, in which the output of one computational block is passed to the next
one sequentially. This is in sharp contrast to the organization of the primate
visual cortex, in which feedback and lateral connections are abundant. In this
work, we propose a computational model for the role of lateral connections in a
given block, in which the weights of the block vary dynamically as a function
of its activations, and the input from the upstream blocks is iteratively
reintroduced. We demonstrate how this novel architectural modification can lead
to sizable gains in performance, when applied to visual action recognition
without pretraining and that it outperforms the literature architectures with
recurrent feedback processing on ImageNet.
‹Figure 1: The generic way to modify an existing neural network. (a) The original feed forward network with N convolutional blocks, each containing one or more layers. (b) The application of our method: the last R convolutional blocks are augmented with recurrent connections. At each recurrent iteration, the output of block N − R is passed to the first recurrent block. Once the K iterations of the first recurrent block are completed, this block’s output is passed to the second recurrent block and so on. (Related Work)›