Karol
Sajnok
Centre for Theoretical Physics PAN
Abstract: Physical backpropagation aims to train physical processors by using their own dynamics to evaluate gradients, avoiding the need for an accurate numerical model. I will present two approaches developed for classical and quantum systems. Near-Equilibrium Propagation (NEP) extends equilibrium-based learning to nonlinear, driven-dissipative wave systems, using the response to a weak perturbation to compute local gradients. Its application to exciton-polariton condensates has been investigated both numerically and experimentally. Analog Quantum Rewind Training (AQuReT) uses a cost-dependent perturbation followed by reversed unitary evolution to extract gradients from measurements of local quantum observables. I will discuss its advantages over conventional quantum training methods under limited computational budgets, along with the theoretical foundations, experimental implementations, and limitations of both approaches.