Thermodynamic-inspired computing with oscillatory neural networks

Descripción del proyecto

En una investigación se explora cómo la física podría hacer el trabajo pesado en las redes neuronales

El consumo energético creciente de las ingenierías informáticas se ha convertido en una preocupación medioambiental, a pesar de los avances en la tecnología de los semiconductores. Las arquitecturas informáticas actuales siguen siendo ineficientes desde el punto de vista energético, están diseñadas principalmente para tareas específicas y son vulnerables al ruido, el calor y la variabilidad. El equipo del proyecto THERMODON, financiado por el CEI, adopta el concepto de «dejar que la física se encargue de la informática» utilizando el ruido, el calor y las variabilidades para lograr una informática eficiente desde el punto de vista energético. La investigación consiste en integrar la termodinámica con la computación neuromórfica para desarrollar una arquitectura innovadora capaz de computación termodinámica y autoorganización. Aplicando los principios termodinámicos a las redes neuronales oscilatorias, los investigadores desarrollarán un paradigma informático novedoso que transforma la arquitectura informática en un sistema dinámico y autoorganizado que interactúa abiertamente con el entorno.

Objetivo

There is a pressing need to address the power consumption of computing, which keeps rising to the point it has become an environmental concern. Despite the remarkable progress in semiconductor technology, computing architectures are still energy inefficient, engineered for deterministic tasks as well as susceptible to noise, heat, and variations. Instead of massively over-designing architectures to compute with an acceptable degree of reliability, this research aims to “let physics do the computing” and harness noise, heat and variabilities for energy efficient computing.

At the heart of the proposed paradigm is the thermodynamics of open systems entwined with neuromorphic computing. THERMODON aims to develop an unconventional neuromorphic architecture to thermodynamically compute and self-organize (“learn”). I hypothesize that the natural thermodynamics of appropriately engineered architecture can harness noise, heat, and variations to self-organize toward energy efficient “solutions” to “problems” posed by external potentials. I will develop such architecture with neuromorphic oscillatory neural networks that I master in my lab. This research aims to address how thermodynamic principles can be applied to oscillatory neural networks to derive learning rules that are unsupervised, continuously adapting and transforming the architecture into a dynamic “self-organizing” and “open interactive” system that learns, infers and interacts with the environment.

THERMODON will bring breakthrough innovations in thermodynamic computing models and AI-specialized hardware to enable online training and inference to intelligent systems. The interdisciplinary research in this project between neuromorphic computing and thermodynamics opens a new and exciting area in computer architecture, triggering a paradigm shift in edge AI computing as well as an immediate impact as a hardware accelerator platform.