Learning in the Machine: To Share or Not to Share?

TL;DR

This paper investigates whether weight-sharing in neural networks is essential by exploring free convolutional networks that relax this assumption, showing they can perform comparably to traditional models in vision tasks.

Contribution

It introduces Free Convolutional Networks, demonstrating that weight-sharing is not necessary for effective learning in computer vision, especially with translationally augmented data.

Findings

Free Convolutional Networks match standard architectures in performance.

Weight-sharing is a pragmatic but not essential optimization.

Translationally augmented data enables learning of invariant representations.

Abstract

Weight-sharing is one of the pillars behind Convolutional Neural Networks and their successes. However, in physical neural systems such as the brain, weight-sharing is implausible. This discrepancy raises the fundamental question of whether weight-sharing is necessary. If so, to which degree of precision? If not, what are the alternatives? The goal of this study is to investigate these questions, primarily through simulations where the weight-sharing assumption is relaxed. Taking inspiration from neural circuitry, we explore the use of Free Convolutional Networks and neurons with variable connection patterns. Using Free Convolutional Networks, we show that while weight-sharing is a pragmatic optimization approach, it is not a necessity in computer vision applications. Furthermore, Free Convolutional Networks match the performance observed in standard architectures when trained using properly translated data (akin to video). Under the assumption of translationally augmented data, Free Convolutional Networks learn translationally invariant representations that yield an approximate form of weight sharing.