Talk by Alexander Dimitrov (INM-6/IAS-6 Seminar)
Host: Markus Diesmann
Neural tissue digital twins in Intel’s Loihi, and emerging mathematical problems
We see a natural match between hardware realizations of data-based neuronal models produced by the neuroscience research endeavor and the neuromorphic hardware abstractions forming the foundation of current and future neuromorphic chips. It seems we are approaching the stage when we can design true biomimetic neuromorphic systems which can essentially replicate a functioning brain tissue in hardware. Our long-term goal is to create simulations of neural tissue that can be used in biomedical research and clinical settings, using modern neuromorphic hardware. Here we report on the initial steps in this process. In this presentation we focus on Intel’s Loihi architecture, as it is at present one of the most powerful neuromorphic platforms with specialized digital hardware and significant software support. To use that architecture, we developed a mapping method from classical GLIF point models to Loihi realizations. We validated the Loihi realization against classic simulations and quantified the differences in state space evolution. We also investigated the sensitivity of both implementation against one another, leading to the more precise definition of neural simulation equivalence classes in different simulation engines and computational paradigms.
The initial investigation lead to a number of open questions along which we are undertaking further investigations. We discuss several:
- What are the appropriate cost functions and comparison scales for validating neural simulations?
- Are neuromorphic simulations compatible with data-based models?
- What optimization tools can be used to achieve closer approximations of classical simulations in neuromorphic hardware?
Departments of Mathematics and Statistics
Washington State University Vancouver