Deep Convolutional Neural Network Design Patterns

TL;DR

This paper analyzes recent deep learning architectures to identify core design principles and introduces innovative network structures like FractalNet, Stagewise Boosting Networks, and Taylor Series Networks, aiming to guide practitioners and inspire future research.

Contribution

It synthesizes collective knowledge to uncover underlying principles for CNN design and presents novel architectural patterns with available implementation code.

Findings

Identified key principles underlying modern CNN architectures.

Proposed new network structures: FractalNet, Stagewise Boosting Networks, Taylor Series Networks.

Provided open-source code for the proposed architectures.

Abstract

Recent research in the deep learning field has produced a plethora of new architectures. At the same time, a growing number of groups are applying deep learning to new applications. Some of these groups are likely to be composed of inexperienced deep learning practitioners who are baffled by the dizzying array of architecture choices and therefore opt to use an older architecture (i.e., Alexnet). Here we attempt to bridge this gap by mining the collective knowledge contained in recent deep learning research to discover underlying principles for designing neural network architectures. In addition, we describe several architectural innovations, including Fractal of FractalNet network, Stagewise Boosting Networks, and Taylor Series Networks (our Caffe code and prototxt files is available at https://github.com/iPhysicist/CNNDesignPatterns). We hope others are inspired to build on our preliminary work.