Classical and quantum physical systems that evolve according to an Ising Hamiltonian are currently attracting broad research interest as novel computing architectures for solving optimization problems that cannot be tackled efficiently on large scales by conventional electronics. However, all the proposed settings either involve a limited number of spins or lack of scalability.
The proposal aims at developing a novel class of photonic Ising machines that exploits the spatial degree of freedom of light for parallel processing of a vast number of spins with programmable couplings. Employing spatial light modulation technology in neural network schemes, the project intends to demonstrate a major breakthrough in the realization of high-speed and scalable novel computing hardware for hard optimization problems and machine learning implementations.
The approach is based on the idea that several thousands of spin variables can be encoded into a phase and amplitude spatial modulation of the optical wavefront. Using designed optical wave-mixing devices that couple the optical spins by linear interference, light propagation is tailored by a recurrent, neural network inspired, measurement and feedback method to evolve toward the ground state of the target spin Hamiltonian. Within this scheme - scalable in terms of spin number and practical resources - we address photonic simulation of zero-temperature ground states and phase transition phenomena for various classical spin models and solve large-scale complex optimization instances with thousands of nodes.
This multidisciplinary project combines paradigms from optics, statistical mechanics, complex systems and optimization algorithms into a cutting-edge spatial photonic setup that will lay the grounds for large-scale all-optical computing.
Spatial light modulation and optics in complex media
Recent years witnessed an enormous amount of results concerning the controlled transmission of light in multiple-scattering random media. An optimal spatial shaping of the optical wavefront phase enables focusing and imaging through disordered and biological samples, as well as applications in information processing ranging from ultracompact spectrometers, metasurfaces, and multimode fibers, to cryptography and universal gates. The PI recently demonstrated phase and amplitude light modulation for sensing tumor morphodynamics beyond standard microscopy techniques. In general, light interaction with disordered nonlinear media is modeled in terms of transmission matrix products, which suggests that wave-mixing devices can act as spatial photonic processors performing large linear matrix operations. Our idea is to use optical transmission matrices implemented by scattering media and programmable optical masks to map arbitrary Ising models.
How the computer of the future will be? The development of novel computing architectures represents one of the main challenges of modern science, with implications that would be far-reaching. A physical system that can efficiently solve truly large NP-hard problems would drastically impact future technology and, consequently, every aspect of our society. The SPIM proposal has been conceived to open a new avenue in photonic computing. Our idea of using screen pixels as computational units (spins) may represent the basis for future computational devices and artificial intelligence hardware that exploit the fundamental properties enshrined in the wave nature of light: high-speed, low-power, optical passivity, and parallelization.
We are proposing an entirely new approach. Full achievement of the objectives requires an interdisciplinary laboratory that combines cutting-edge experimental photonics with the frontiers of scientific computing. In fact, the SPIM activity outlines several research directions requiring special support and considerable resources. For instance, while in the present proposal the challenge is to develop a photonic platform able to find the ground state of arbitrary Ising model, we do not focus the attention on how fast this solution is found. Further engineering the most recent spatial light modulation technologies with dedicated efforts, the iteration time of our machine can be pushed down to the microsecond time scale. Moreover, using single-photon sources and detectors the whole scheme can be implemented with quantum light, thus opening fascinating possibilities for quantum computing. On the other hand, the pioneering idea of implementing neural network schemes with physical objects can be generalized to solve machine-learning problems from recognition to prediction. The search for these "real neurons" will be a central theme in the new era of artificial intelligence.