Towards a More General Understanding of the Algorithmic Utility of Recurrent Connections
AbstractLateral and recurrent connections are ubiquitous in biological neural circuits. The strong computational abilities of feedforward networks have been extensively studied; on the other hand, while certain roles for lateral and recurrent connections in specific computations have been described, a more complete understanding of the role and advantages of recurrent computations that might explain their prevalence remains an important open challenge. Previous key studies by Minsky and later by Roelfsema argued that the sequential, parallel computations for which recurrent networks are well suited can be highly effective approaches to complex computational problems. Such “tag propagation” algorithms perform repeated, local propagation of information and were introduced in the context of detecting connectedness, a task that is challenging for feedforward networks. Here, we advance the understanding of the utility of lateral and recurrent computation by first performing a large-scale empirical study of neural architectures for the computation of connectedness to explore feedforward solutions more fully and establish robustly the importance of recurrent architectures. In addition, we highlight a tradeoff between computation time and performance and demonstrate hybrid feedforward/recurrent models that perform well even in the presence of varying computational time limitations. We then generalize tag propagation architectures to multiple, interacting propagating tags and demonstrate that these are efficient computational substrates for more general computations by introducing and solving an abstracted biologically inspired decision-making task. More generally, our work clarifies and expands the set of computational tasks that can be solved efficiently by recurrent computation, yielding hypotheses for structure in population activity that may be present in such tasks.Author SummaryLateral and recurrent connections are ubiquitous in biological neural circuits; intriguingly, this stands in contrast to the majority of current-day artificial neural network research which primarily uses feedforward architectures except in the context of temporal sequences. This raises the possibility that part of the difference in computational capabilities between real neural circuits and artificial neural networks is accounted for by the role of recurrent connections, and as a result a more detailed understanding of the computational role played by such connections is of great importance. Making effective comparisons between architectures is a subtle challenge, however, and in this paper we leverage the computational capabilities of large-scale machine learning to robustly explore how differences in architectures affect a network’s ability to learn a task. We first focus on the task of determining whether two pixels are connected in an image which has an elegant and efficient recurrent solution: propagate a connected label or tag along paths. Inspired by this solution, we show that it can be generalized in many ways, including propagating multiple tags at once and changing the computation performed on the result of the propagation. To illustrate these generalizations, we introduce an abstracted decision-making task related to foraging in which an animal must determine whether it can avoid predators in a random environment. Our results shed light on the set of computational tasks that can be solved efficiently by recurrent computation and how these solutions may appear in neural activity.
