Unsupervised learning for robust working memory

Abstract

AbstractWorking memory is a core component of critical cognitive functions such as planning and decision-making. Persistent activity that lasts long after the stimulus offset has been considered a neural substrate for working memory. Attractor dynamics based on network interactions can successfully reproduce such persistent activity. However, it suffers from a fine-tuning of network connectivity, in particular, to form continuous attractors suggested for working memory encoding analog signals. Here, we investigate whether a specific form of synaptic plasticity rules can mitigate such tuning problems in two representative working memory models, namely, rate-coded and location-coded persistent activity. We consider two prominent types of plasticity rules, differential plasticity targeting the slip of instant neural activity and homeostatic plasticity regularizing the long-term average of activity, both of which have been proposed to fine-tune the weights in an unsupervised manner. Consistent with the findings of previous works, differential plasticity alone was enough to recover a graded-level persistent activity with less sensitivity to learning parameters. However, for the maintenance of spatially structured persistent activity, differential plasticity could recover persistent activity, but its pattern can be irregular for different stimulus locations. On the other hand, homeostatic plasticity shows a robust recovery of smooth spatial patterns under particular types of synaptic perturbations, such as perturbations in incoming synapses onto the entire or local populations, while it was not effective against perturbations in outgoing synapses from local populations. Instead, combining it with differential plasticity recovers location-coded persistent activity for a broader range of perturbations, suggesting compensation between two plasticity rules.Author SummaryWhile external error and reward signals are essential for supervised and reinforcement learning, they are not always available. For example, when an animal holds a piece of information in mind for a short delay period in the absence of the original stimulus, it cannot generate an error signal by comparing its memory representation with the stimulus. Thus, it might be helpful to utilize an internal signal to guide learning. Here, we investigate the role of unsupervised learning for working memory maintenance, which acts during the delay period without external inputs. We consider two prominent classes of learning rules, namely, differential plasticity, which targets the slip of instant neural activity, and homeostatic plasticity, which regularizes the long-term average of activity. The two learning rules have been proposed to fine-tune the synaptic weights without external teaching signals. Here, by comparing their performance under various types of network perturbations, we reveal the conditions under which each rule can be effective and suggest possible synergy between them.