A team of researchers of the Department of Physics of Sapienza University of Rome and the Italian National Research Council Institute for Complex Systems, proposed a new model of artificial intelligence which uses the propagation of waves to perform complex activities such as understanding a text or driving a car. The results of the study have been published on the Physical Review Letters journal
Artificial intelligence generally evokes the image of a complex computer network, arranged in endless rooms, which process immense amounts of data for the most diverse functions: image recognition, text translation, strategies for defeating a chess champion or optimising stock portfolios.
However, the human superiority over the machine is still evident if we analyse the energy consumption needed to train a computer to recognise, for example, a cat from a dog, and then compare it to the few teaspoons of sugar that are enough for our brain to do much more complex operations, such as understanding a text or driving a car. The gap between humans and machine is merciless: it is estimated that “teaching” a neural network to perform complex activities pollutes as much as 20 times a human being in a year.
In the last decade, studies in this field have focused on a new model of artificial intelligence called “neuromorphic”, because it is similar in function to the human brain and as such with potentially very low energy consumption for learning.
It is in this field that Giulia Marcucci, Davide Pierangeli, and Claudio Conti of the Department of Physics of Sapienza and the National Research Council Institute for Complex Systems, proposed a completely innovative model to create an artificial intelligence system which, by exploiting the propagation of light waves, can compute in an “economical” way.
The paper, published on the Physical Review Letters journal, is based on the fact that chemical and electrical waves, commonly referred to as brain waves, propagate in our brain, allowing us to perform the complex operations underlying our behaviour. Just like newborn children can move their legs, use their eyes to explore the environment, close or open their mouth to react to stimuli, or even swim, without their brain ever having been instructed to do so, the physical system studied by the team is capable of performing complex operations without training.
“In our work – says Claudio Conti – we have shown how, by exploiting the propagation of light waves of a laser beam, and their interaction in a crystal, we can solve different types of computational problems in a potentially very energy-efficient way. Applications include big data classification, image analysis and new neuromorphic calculators.”
Traditional artificial intelligence involves the creation of electronic systems to be trained in every aspect with a very high expenditure of energy. This study has shown that waves have a computational capacity with very low energy consumption, whose advantage is evident, especially in relation to pollution and environmental impact. The next step will be to carry out experimental checks on the system in order to replace the complex and expensive computer networks with small optical devices that use the intelligence of the waves.
Theory of Neuromorphic Computing by Waves: Machine Learning by Rogue Waves, Dispersive Shocks, and Solitons – G. Marcucci, D. Pierangeli, C. Conti. – Physical Review Letters 2020, 125, 093901. DOI: https://doi.org/10.1103/PhysRevLett.125.093901
Department of Physics Sapienza University of Rome
Source Sapienza University of Rome