Creating microprocessors capable of replicating biological learning systems, to make artificial intelligence more flexible, efficient, and environmentally sustainable is the challenge launched by an international group of researchers coordinated by the Neuromorphic AI Lab (NUAI Lab) at the UTSA (University of Texas at San Antonio) - which includes Vincenzo Lomonaco, one of Italy’s leading experts in Continual Learning, a researcher at the Department of Computer Science at the University of Pisa and one of the authors of the article “Design principles for lifelong learning AI accelerators”, recently published in the prestigious scientific journal Nature Electronics.
“The fallibility of Artificial Intelligence is still too high, and this is because AI, as we know it today, is based on non-adaptable machine learning systems, which make it incapable of dealing with new conditions not previously encountered during the training process, explains Vincenzo Lomonaco. “In fact, we make it learn a large amount of information all at once, but if something new emerges on a certain topic, we have to update the system from scratch. In addition to being inefficient, AI has a costly economic and environmental impact, considering the high energy consumption and consequent CO2 emissions of this process”.
Vincenzo Lomonaco
Upgrading an AI system can cost up to several million euros. Furthermore, according to a recent study carried out by the University of Massachusetts, training several large AI models can emit five times the amount of carbon dioxide emitted by an average American car during its life cycle, including the manufacturing process.
One solution, according to Lomonaco and the other researchers of the Neuromorphic AI Lab - coordinated by Professor Dhireesha Kudithipudi -, is represented by Continuous Automatic Learning (also known as Continual Learning or Lifelong Learning), which would allow AI to assimilate a large amount of knowledge in sequence, without forgetting what has been learned previously.
“To realise such a learning system, it is necessary to change the current computational paradigms, eliminating the current infrastructural constraints,” continues Lomonaco, “that is why, with some colleagues at the NUAI Lab in San Antonio, we laid the groundwork for a new incremental learning system, based on hardware-software co-design. Designing hardware and software components simultaneously, to create a robust and autonomous lifelong learning system for AI. All based on next-generation algorithms that, working like human intelligence, allow AI to increase its knowledge continuously, faster, and more efficiently, with energy consumption close to that of a light bulb”.