Probability theory from neurons to cognition
Abstract
The core objective of this proposal is to demonstrate how Hilbert space representations can be used to derive detailed neural circuits. We hypothesize that the latent state represented by populations of spiking neurons is the space of points in high-dimensional Hilbert spaces. We will test the neural plausibility of circuits generated under this hypothesis by comparing their structure and performance to biological data. We will frame the work by taking specific behavioral tasks and their associated brain regions and constructing functional neural models of these regions. We will build upon our prior success in constructing large-scale functional brain models, while deploying latent Hilbert space representations. This approach allows us to formulate both novel and commonly understood models of brain area function as probabilistic models. Our approach applies to traditionally cognitive and traditionally non-cognitive behavior. Consequently, our approach is general and, if correct, will provide a parsimonious explanation for neural representation and computation across a wide range of brain areas and levels of analysis. The primary impact of this work will be to establish the claim that uncertainty in neural representations can be best modeled as Hilbert space representations and operations. Doing so will provide a new explanatory framework for understanding how biological organisms act in noisy or uncertain natural environments that captures mechanisms spanning from neurons to cognition. We will use the framework to make detailed predictions regarding a variety of experimentally testable properties, including neural dynamics, tuning curves, spike statistics, behavioral response distributions and timing, and so on. As well, we expect to discover and characterize novel algorithms that can be extracted from our description of neural mechanisms. Such algorithms will improve current state-of-the-art in machine learning and AI, and provide networks that can be implemented on standard accelerators or nascent neuromorphic computing platforms.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 07, 2024
- Source ID
- FA95502310644
Entities
People
- Chris Eliasmith
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Waterloo