Dan Greenberg
Sep 2, 2018 · 1 min read

This is a good explanation of CNNs specifically. We would augment it a bit by looking at a higher level of architecture. What you find there is that biology has evolved the idea of a pre-processing stage (which varies by the type of signal to be processed) multiple times. And, in our field of human text processing, we believe a powerful enough pre-processor enables simple structured machine learning at later stages, obviating the need for deep learning (and its heavy data and processing demands). See https://www.textician.com/onrepresentation/ (as well as others in our blog). In vision processing specifically, see https://arxiv.org/ftp/arxiv/papers/1801/1801.00631.pdf from Gary Marcus and https://arxiv.org/pdf/1806.06850.pdf (regarding polynomial methods).

    Dan Greenberg

    Written by

    Seeking interesting business, and businesses, to develop. Please forgive occasional snark.