Cognitive computations are accomplished by a wide variety of complex activity dynamics in the networks of the brain. We focused specifically on the neural dynamics responsible for working memory and decision making. Working memory is the process by which pieces of information are stored for a short period of time and used to perform some calculation or computation. Decision making requires the accumulation of evidence to determine the best choice between multiple alternatives. We studied these phenomena by building neuronal network models which generated spatiotemporal activity to encode these cognitive processes.

Intellectual Merit: First, we showed how the multilayered structure of cortex is important for making these computations robust. One common working memory task developed for test animals is that of a spatial navigation task in which the subject must remember their location along a linear track. We showed that encoding position with a neural activity bump that spans multiple layers of a network can reduce encoding error that might arise due to internal noise or network variability. Moreover, we showed that such position encoding networks can be utilized to store information about previously visited locations, not just an animal's current location, when neural activity propagates as a moving front. This work also prompted further study for conditions of wave propagation in neural activity models, providing a fundamental mathematical framework for understanding propagation thresholds in nonlocal models. The idea of modeling memory for previously visited locations also springboarded additional work studying optimal strategies for animals foraging in an environment that they slowly deplete. This study pushed forward new theory on the statistics of first passage times in discrete models.

A second aim of the project was to understand how the dynamics of the synaptic connections of neural networks may shapes information storage. A key study focusing on this aim explored how short term facilitation, the process by which synapses of frequently active neurons are strengthened on the timescale of seconds, may contribute to the interaction of multiple items stored in memory over time. Interference is a commonly observed phenomenon in working memory experiments whereby previously remembered items can bias one's memory for currently stored items. We found that short term facilitation can account for this in a neuronal network model, as the facilitated synapses from previously remembered items can attract the neural activity representing the currently remembered item. Our model made a number of experimentally testable predictions concerning the behavioral responses of subjects performing back-to-back working memory trials, and potentially the neural activity recordings to be expected in non-human primates. We are currently in the process of working with experimental collaborators to test these predictions in collected behavioral data and neural recordings.

In related work, we have explored how neural activity interactions within a single trial might account for working memory item interference of simultaneously stored items. Indeed, we found that the distance between items in feature space can strongly shape the nature of item interference, in the context of a neural network model. These predictions are also being validated currently with collected behavioral data from collaborators.

We also studied how multiple layers of a network could implement adaptive decision making. A key finding was that the rate of environmental changes can be encoded by the strength of synapses between the layers which is tuned by long term plasticity. This suggests a fundamental neural mechanism for adaptive evidence integration in networks, which we are exploring currently in a subsequent funded project.

Broader impacts: These results were disseminated broadly through papers (9) in refereed applied math and neuroscience journals; conference talks and seminars; and tutorials. One tutorial was given at the International Conference on Mathematical Neuroscience in June 2019 to 60 students from all over the world, about half of whom were women. One PhD student, Daniel Poll, graduate in May 2017, and is currently a postdoc at Northwestern. Another PhD student, Kate Nguyen, will likely graduate in August 2020. Three undergraduates were funded, leading to two publications. At present, the PI is training two female students from countries underrepresented in science, a female undergraduate, and a female postdoc, so more than half of the trainees in the group are from underrepresented groups in STEM.

Last Modified: 11/12/2019
Modified by: Zachary P Kilpatrick