Physics-inspired (π-) computing

Natalia Berloff is a professor of applied mathematics at the Department of Applied Mathematics and Theoretical Physics (DAMTP), University of Cambridge and Fellow of Jesus College, Cambridge. She gained her undergraduate qualifications the Lomonosov Moscow State University and PhD from Florida State University. Before her appointment as a faculty member in DAMTP in 2002, she had been the beneficiary of the UC Presidents Postdoctoral Fellowship and subsequently held the post of assistant professor in the Program in Computing at the University of California, Los Angeles. Her areas of research include nonlinear waves, quantum fluids, complexity, and analogue computing.

Given the colossal reliance on solving optimisation problems in industries such as the logistical, pharmaceutical, and data management sectors, it is increasingly important that quick and novel techniques are found that can accurately determine optimal solutions. I work on analysing and benchmarking physical platforms that efficiently solve NP-hard optimisation problems. Such platforms operate by locating the ground state of specially constructed spin Hamiltonians.

My research focuses on enhancing the efficiency of physics-inspired optimization methods. Many of those methods rely on the dynamics of large networks of interacting oscillators.  I aim to improve our understanding of the intricate interplay between the complex dynamics of these networks and the sophisticated structure inherent to hard optimization problems. To achieve this, I employ numerical simulations as well as analytical calculations.

My research mainly involves the use of physical systems, especially polariton condensate network, to realise unconventional computing paradigms, such as neuromorphic and reservoir computing. These computing approaches based on a network of nodes demonstrate structural similarity to many physical systems. Through numerical simulations guided by theoretical calculations, my objective is to find ways to utilise these physical systems to implement these computing architectures and solve difficult optimisation problems traditionally tackled by other methods, such as deep neural networks.

I am interested in physically inspired computation and solving NP hard optimization problems. By exploiting the physical representations of the computational problems, we can analyze the physical networks geometry, dynamics of topological defects, landscape roughness and connect them to the computational complexity. By developing an analytical description and simulating the behavior of these physical systems, I am trying to find approaches that will allow the system to avoid the local minima to find more reliable solutions faster.