Roadmap for Implementation Optical Extreme Learning Machines

Abstract

The principles of machine learning are discussed in terms of non-linear mappings and signal mixing. We considered the characteristics of the capabilities of algorithmic reservoir computers based on the software implementation of artificial neurons with random weights of input signals, and physical reservoir computers using various random and non-linear phenomena. The main elements of the concept of extremal learning machines are outlined, their features and the algorithm for learning a linear output with comb regularization by the method of pseudo-inverse Moore-Penrose matrices are described. The study found the subdivision of optical reservoir computers by types and spatial scales of the physical processes used. Optical Extreme Learning Machines (OELM) are identified as a promising area of implementation. The structure of the diffractive OELM, its principle of operation based on the scattering by random elements, and its inherent limitations are described. The use of plasmonic metal nanostructures to reduce the size of such processors is proposed. To implement the quantum version of the OELM signal with frequency modulation, a new approach to processors operating on systems of artificial atoms with random interactions has been formulated. The study proposed to incite and read signals using the methods of three-pulse femtosecond pump-probe spectroscopy. The need for a low level of relaxation during the processor cycle is noted for stable operation of the circuit. The study discussed an individual atom generating high harmonics in a strong laser field as the fundamental quantum limit of the processor, and described the physical mechanism providing this effect. For each of the three OELM options, critically important tasks are formulated, the consistent solution of which will significantly bring the creation of technologically significant OELM closer.