Project

Back to overview

Neuromorphic Algorithms based on Relational Networks

Applicant Cook Matthew
Number 182539
Funding scheme Project funding
Research institution Institut für Neuroinformatik Universität Zürich Irchel und ETH Zürich
Institution of higher education University of Zurich - ZH
Main discipline Information Technology
Start/End 01.07.2019 - 30.06.2023
Approved amount 1'200'000.00
Show all

Keywords (3)

neuromorphic computing; neuromorphic engineering; neuromorphic algorithms

Lay Summary (French)

Lead
Ces dernières années ont vu une explosion du nombre d'architectures dites neuromorphiques. Celles-ci, bien que conformes à la vision de Carver Mead de répliquer la biophysique d'un neurone par la biophysique d’un transistor opérant dans le domaine analogique, ont plus largement fait émerger l'idée de supporter de nouveaux algorithmes bio-inspirés. Cependant, à ce jour, seuls peu de ces algorithmes peuvent être implémentés efficacement sur ces dernières. Notre projet vise a étudier les réseaux relationnels. Dans ceux-ci, la structure du calcul est donnée par la topologie du réseau: un graphe dans lequel la donnée est transmise suivant ses arêtes, les calculs étant effectués dans les nœuds qui la reçoive et la retransmette. Différents formalismes déjà existant servent d’exemples d'algorithmes relationnels, mais il existe de nombreuses directions restant à explorer, tant de manière théorique ou applicative que dans leur pratique et leur implémentation sur des plateformes neuromorphiques.
Lay summary

Contenu et objectifs du travail de recherche
Notre principal objectif est le développement de nouvelles techniques qui étendent notre compréhension du type de calcul qui peut être supporté par les réseaux relationnels ainsi que leur application. 

Nous demontrerons le large potentiel des algorithmes basés sur des réseaux relationels dans diverses domaines d'application tels que l'apprentissage de séquences pour des réseaux de neurones à spikes, l'apprentissage profond dans des réseaux relationnels, la fusion de senseurs, l'odométrie visuelle, la coordination d'action, leur sélection ainsi que leur apprentissage dans des graphes de controlleurs. De plus, nous souhaitons montrer leur compatibilité avec les architectures neuromorphiques en démontrant leur implémentation sur ces dernières.

Contexte scientifique et social du projet de recherche
De part la pratique des algorithmes basés sur des réseaux relationnels: dans leur conception et leur implémentation, nous montrerons qu'ils constituent une alternative intéressante et efficace vis a vis d'algorithmes de calculs conventionnels. 

Keywords
Réseaux de relations, Algorithmes neuromorphiques, Calcul sur architecture non von-Neumann.

 
Direct link to Lay Summary Last update: 04.10.2019

Responsible applicant and co-applicants

Employees

Abstract

Recent years have seen a rapid expansion of neuromorphic computing platforms. In the last decade there has been an explosion of neuromorphic hardware platforms, and even the term ``neuromorphic'' has broadened significantly in usage, to encompass not just Mead's original vision of analog transistor circuits that closely model the behavior of a neuron, but more generally any form of computing hardware which is designed to support some sort of neural-like computation.Similarly, the term ``neuromorphic algorithms'' has grown from its original meaning of ``algorithms that use brain-like neural networks to replicate specific brain functions'' to a more broad meaning of ``algorithms that have a biologically-inspired architecture and achieve some sort of brain-inspired behavior.'' However, the availability of neuromorphic algorithms currently lags well behind the advances in available hardware.However, these platforms are not yet widely used. There was a general expectation that these neuromorphic hardware systems would fill a need for the people who do software simulations of neural networks. Such simulations run relatively slowly on standard computers, and they are not easily sped up with multiple cores or GPUs due to the need for constant inter-core communication. Specialized hardware would in principle provide dramatic acceleration for these simulations, but in practice such hardware has not yet become common among neural network researchers.Algorithms based on relational networks are well suited to neuromorphic hardware. In our research group, we have created a number of complex algorithms ranging from navigation to optic flow detection and sensor fusion, targeted at neuromorphic hardware platforms. In some cases this targeting is direct, involving a low-level implementation directly on the neuromorphic hardware, while in other cases the targeting is more abstract, where we write software using primitives that could easily be mapped into a neural-network form at some point in the future, if needed. (And in some cases this mapping has indeed taken place, bringing high level algorithms into a form that can be implemented on neuromorphic devices.) The higher level allows investigation of large systems which, if implemented at a neural level, might need more neurons than are available in current hardware systems.When we look at the form of these algorithms, we see that they typically have a completely different kind of design from traditional algorithms. We call them "relational-network-based algorithms", or just "relational algorithms". In these algorithms, the structure of the computation itself is given by a network of relations: Data flows along the edges of the network, and simple calculations happen at the nodes of the network, so the computation is fully distributed with no global control.Specific new directions for relational algorithms:Many existing formalisms are instances of relational algorithms. However, there are a large number of compelling directions waiting to be explored. These include spiking sequence learning, spatial networks, learning in deep relational networks, uncertainty, sensor fusion and algorithm fusion, computing on non-grid graphs, visual ego-motion, action coordination, optimizing action selection, action learning, active sensing, putting large networks onto hardware, developing compilation tools for relational algorithms targeting neuromorphic hardware, and developing specifications for hardware to support relational algorithms.This proposal aims to explore all of these directions, as described in the following pages.The goal:In addition to directly yielding powerful, widely applicable algorithms, we expect our developments to provide new techniques that will effectively become new language features and design approaches for relational-processing languages.In short, the overall goal of this proposal is to massively extend the types of computations that can be done by relational algorithms. By making them much more broadly applicable, as well as being efficiently implementable on neuromorphic hardware, we expect them to become a very attractive alternative to traditional computing.
-