Applying a Biologically Inspired Software Retina to Convolutional Neural Networks

A fundamental issue with convolutional neural networks is that there never is enough memory or computational power for them to operate directly on large, high resolution images. The very same challenge is present in primate vision, and over the course of evolution it has been solved with variable resolution foveated imaging. Taking reference from nature, this research proposes to improve on convolutional neural networks by reducing their memory requirements and training times, as well as making the networks more robust to variances in object scale and rotation. It also aims to make neural networks more available to non-desktop architectures. This will be achieved with the help of a biologically motivated software retina that subsamples from the original image to cull redundant information and produce a compressed image. The proposed works will be an expansion on my current research into the software retina