Professor Alexander Dimitrov, from Washington State University (WSA) will give the talk: Neuromorphic computing systems: what do we do, what can we do, and how to do it?

Day: 31 / 07 / 2024
Time: 14h00
Online via http://aovivo.ic.unicamp.br

Abstract

We are at the crossroads of powerful currents in science and technology, spurred by recent advances in both neuroscience research, and neuromorphic engineering. A recent focus on brain studies has produced a wealth of new, multimodal structural and functional neural data. At the same time, advances in electronics and the search for unconventional computational paradigms led to the creation of neuromorphic systems like Intel's Loihi, IBM's NorthPole, and the EU SpiNNaker and BrainScaleS. 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. We also see a new paradigm in computing abstractions, possibly advancing beyond the von Neumann paradigm, and through Moore's scaling law.

One major unresolved question is the practical issue of programming neuromorphic systems. How can we efficiently combine neuromorphic modules to achieve specific dynamics, and later – specific task by that dynamics? As with the original development of computers, tools from applied mathematics, with 70-80 years additional development past the complement of the 1940-s will be crucial for this: nonlinear optimization, including evolutionary programming; linear and nonlinear control; sensitivity analysis; neural-derived cost functions. All neuromorphic simulations should be validated against classical numeric algorithms, eg using neuroinformatics tools developed for neural data and model validation.

There is also a broader question: what is computable with these new systems? Neuromorphic system elements have their own dynamics, which is steerable to a degree. How close can such modules approximate dynamics of interest to us? What are appropriate cost functions to assess levels of approximation? And in our context: to what degree can such a neural dynamical system be represented in current neuromorphic computing environment?

Professor Dimitrov will report on his lab's work with Loihi along these questions.

Summary

We are at the crossroads of powerful currents in science and technology, driven by recent advances in both neuroscience research and neuromorphic engineering. A recent focus on brain studies has produced a wealth of new multimodal structural and functional neural data. At the same time, advances in electronics and the search for unconventional computing paradigms led to the creation of neuromorphic systems such as Intel's Loihi, IBM's NorthPole and the European projects SpiNNaker and BrainScaleS. We see a natural correspondence between the hardware realizations of data-driven neuronal models produced by neuroscience research and the neuromorphic hardware abstractions that form the basis of current and future neuromorphic chips. We also observe a new paradigm in computational abstractions, possibly moving beyond the von Neumann paradigm and through Moore's scaling law.

An important and unresolved issue is the practical problem of programming neuromorphic systems. How can we efficiently combine neuromorphic modules to achieve specific dynamics and subsequently specific tasks through these dynamics? As with the original development of computers, tools from applied mathematics, with an additional 70-80 years of development since the 1940s complement, will be crucial to this: nonlinear optimization, including evolutionary programming; linear and nonlinear control; sensitivity analysis; cost functions derived from neurons. All neuromorphic simulations must be validated against classical numerical algorithms, for example using neuroinformatics tools developed for data validation and neural models.

There is also a broader question: what is computable with these new systems? The elements of the neuromorphic system have their own dynamics, which can be directed to a certain extent. How closely can these modules approximate dynamics of interest to us? What are the appropriate cost functions for evaluating approximation levels? And in our context: to what extent can a neural dynamical system be represented in today's neuromorphic computing environment?

Professor Dimitrov will report on his laboratory's work with Loihi on these questions.