Learning in the Machine: To Share or Not to Share?
This addresses a fundamental discrepancy between artificial and biological neural systems, offering insights for more biologically plausible AI models, though it appears incremental in scope.
This study investigates whether weight-sharing, a key feature in Convolutional Neural Networks, is necessary for computer vision by exploring Free Convolutional Networks with variable connection patterns. It finds that weight-sharing is not essential, as Free Convolutional Networks achieve performance matching standard architectures when trained on translationally augmented data.
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.