Project description
Brain-like neural networks for energy-efficient AI solutions
The field of narrow AI is advancing, yet there is an increasing demand for more universal approaches to AI capable of operating across a wider range of applications. The EU-funded EXTRA-BRAIN project aims to develop flexible, hardware-friendly, and energy-efficient AI solutions based on brain-like neural networks. These networks aim to overcome the limitations of current AI methods, such as bounded reliability and excessive data dependence. The project leverages insights from computational neuroscience regarding the brain’s information processing principles, learning schemes, and neuroanatomical structures to design these networks. Supporting the models are data optimisation pipelines and an explainability framework. The project will validate this framework across various use cases with different hardware demands.
Objective
In parallel to the current developments in the so-called narrow artificial intelligence (AI) realm, there is an urgent demand for more universal, general AI approaches that can operate across a wider spectrum of application domains with varying data characteristics. It is expected that the emerging sustainable AI methods can be efficiently deployed in the edge-cloud continuum on different hardware platforms and computing infrastructure depending on the real-world task scenarios and constraints including the limited energy budget. In response to this growing demand and emerging trends we propose to adopt a brain-like approach to AI system design due to its promising potential for functional flexibility, hardware friendliness as well as energy efficiency among others. To this end, EXTRA-BRAIN is aimed at developing a new generation of AI solutions based on brain-like neural networks that enable us to overcome key limitations of the current state-of-the-art methods, exemplified by deep learning, such as limited cross-task generalisation and extrapolation to novel domains (bounded reliability), excessive dependence on costly annotated data as well as extensive training and validation processes with heavy demand for compute resources at high energy cost, to name a few. The core brain-like neural network design in our approach derives from the accumulated computational neuroscience insights into the brain's working principles of information processing, key learning schemes and neuroanatomical structures that underlie the brain's perceptual/cognitive phenomena and its functional flexibility. Furthermore, these novel models are supported by data optimisation pipelines, which improve data quality, security and reduce the costs of assembling suitable training data, and an explainability framework to empower the human user. The proposed EXTRA-BRAIN framework will be examined in a diverse set of use cases with different hardware demands in the edge-cloud continuum.
Fields of science
Not validated
Not validated
- natural sciencescomputer and information sciencesartificial intelligencemachine learningdeep learning
- natural sciencesbiological sciencesneurobiologycomputational neuroscience
- natural sciencescomputer and information sciencesdata sciencedata processing
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Keywords
Programme(s)
Funding Scheme
HORIZON-RIA - HORIZON Research and Innovation ActionsCoordinator
100 44 Stockholm
Sweden