Prefrontal circuits underlying working memory encoding and maintenance

PhD candidate: Nicolás Pollán Hauer

Academic tutor: Lluís Alsedà i Soler

Thesis supervisor: Klaus Wimmer

Working memory, the capacity to maintain and manipulate information in our minds when it is no longer available in the environment, is a central function of cognition. One of the most important neuronal correlates of this cognitive function are the so-called persistent neurons, which respond selectively to sensory stimulation and sustain their increased activity even after removing the stimulus. This phenomenon, most frequently observed in the prefrontal cortex, has been successfully described by neural network models with attractor dynamics. However, only a few of the neurons engaged in working memory tasks have persistent activity. Moreover, analysis of the experimental recordings at the population level reveals that the code undergoes a change between the stimulus presentation and the maintenance epochs, which is not compatible with a working memory code that would only rely on stably active persistent cells. The prevalence of this finding has motivated the proposal of alternative mechanisms, but current computational models that explain dynamics fail to include stable epochs or do not provide a clear mechanistic interpretation.

In this thesis, we use statistical data analysis and neural network modeling to investigate whether specialized neuronal subpopulations underlie the stable and dynamic working memory codes.

First, we investigated the connection between the observed dynamics in the working memory code and the functional structure of the prefrontal circuits. We analyzed prefrontal recordings from behaving macaque monkeys and observed that feature selectivity is non-randomly distributed across the neurons. This non-random or structured feature selectivity distribution is related to functional distinct subpopulations whose contrasting activity explains the dynamic to stable transition in the working memory code.

Second, we developed a computational model that represents three functional subpopulations as attractor networks working on different dynamic regimes. The model illustrates how the population structure, which implies different neurons active at different task epochs, is directly related to the dynamic transition in the code. Furthermore, we show how the three-network architecture can be easily extended to account for additional features, such as ramping activity and variable maintenance periods.

Third, our subpopulation-based networks have the functional advantage of being robust against distracting stimuli. The model captures the experimentally observed vulnerability to distractors presented shortly after stimulus removal. Moreover, it predicts that top-down feedback enhances the overall network’s robustness.

In summary, we show how the presence of functional subpopulations in the prefrontal cortex can be related to the dynamic to stable transition in the working memory code and to an enhanced capacity to filter out distracting stimuli. In conclusion, our work reconciles attractor dynamics with the observed dynamic changes in the code, still suggesting that attractor dynamics are essential for working memory maintenance.