Constructing Spiking Networks as a Foundation for the Theory of Manifolds as Computational Substrate

Join us for a talk by Brian DePasquale, Assistant Professor in the Artificial and Biological Intelligence Lab at Boston University. This talk is part of the Kempner Seminar Series, a research-level seminar series on recent advances in the field.

Abstract: In many organisms, spikes serve as the basic unit of neural communication. A growing consensus is that neural state dynamics confined to restricted trajectories (neural manifolds) serves as a powerful abstraction for understanding computation, and one that provides a bridge between data analysis and theoretical models. However, the field has struggled to understand how neural manifolds relate to the mechanistic elements of neural circuits, such as spikes, leading to discomfort amongst some with this abstraction. In this talk, I will introduce a novel framework for constructing artificial neural networks of spiking model neurons that can embody low-dimensional manifolds within their synaptic connections. This framework provides two important advances. First, it provides a comprehensive approach to constructing biophysically-realistic circuit models using empirical data, which I will illustrate using recordings from the primate motor cortex. Second, it establishes that although neural manifolds are not directly observable, they are as mechanistically ‘real’ as other empirically observed quantities and thus should not be treated with undue suspicion. I will present ongoing work extending this framework to networks that perform multiple computations, in an effort to understand the origin of the ubiquitous observation of trial-to-trial spiking variability. To conclude, I will present ongoing efforts in my new lab to devise machine learning approaches that can meet the increasing complexity of neuroscience data while remaining interpretable and biologically relevant. I will highlight two ongoing efforts: state space models to describe the latent neural dynamics that mediate strategy switching during decision making, and newly developed graph neural network approaches for understanding the structure of neural representation in the olfactory system.