Holding some information in mind – i.e. working memory – is thought to require neural activity that persists during the period the information is held in mind. This idea is based on experiments in which animals are presented with a brief sensory stimulus that they must hold in working memory during a delay period (with no stimulus present) to get a reward (e.g. Funahashi et al, 1989). Some neurons show an increase (or decrease) in firing rate when the stimulus is presented and persist with elevated (or depressed) firing during the delay. On the other hand, neurons may need to be flexible, switching among different possible functions, which would seem to be at odds with persistent activity. Here we explore the connection between criticality, persistent activity, and flexible switching using a computational model and comparison to experimental data.
We originally developed our model to explain two prominent experimental observations that have little to do with working memory: 1) subsets of neurons with strongly anticorrelated activity, and 2) scale-free fluctuations of the activity within these subsets (Jones et al, 2021 biorXiv https://doi.org/10.1101/2021.05.12.443799). In this model, scale-free fluctuations emerge from a winnerless competition between two subsets of neurons, mediated by crossing inhibition. It turned out that this model was very similar to that used to study persistent activity in the context of working memory (e.g. Wang, 2002), but previous work did not account for proximity to criticality. Here we show that our model exhibits the best persistence near criticality, but slightly supercritical. However, flexible switching was maximized closer to criticality. Our model predicted that some subsets of neurons in mPFC might operate near criticality to achieve a combination of flexibility and persistent firing. We tested this prediction on recordings of mPFC neurons in rats doing a working memory task and found that 18% of neurons in mPFC behave in accordance with our prediction.