Effective learning algorithms for physical machine learning (PML)

Develop effective learning algorithms for physical machine learning on physics-based ASICs that can efficiently configure the tunable physical parameters θ governing the hardware transformation y = f_p(x, θ), either by running the learning procedure on a separate processor that controls the hardware or, ideally, by performing the learning within the physical hardware itself (physical learning).

Background

Physical machine learning (PML) seeks to learn computations directly at the hardware level by optimizing the physical parameters of a device so that its input–output transformation matches desired behavior. In this framework, inputs are encoded into a physical system and outputs are obtained by measuring the system after it evolves, with tunable physical parameters controlling the transformation.

Despite its promise, PML faces significant hurdles: optimization landscapes can be challenging (e.g., barren plateaus), and real hardware often deviates from simulations (the sim-to-real gap), making purely simulation-based optimization unreliable. Consequently, there is a need for learning algorithms that can configure hardware parameters efficiently, either via an external controller or directly on the hardware itself through physical learning.

References

As a consequence of these challenges, an important open challenge for the subfield of PML is to develop effective learning algorithms that can be used either within a separate processor to configure $\vec{\theta}$ efficiently, or -- ideally -- to use the physical hardware itself for this purpose, i.e., physical learning .

— Solving the compute crisis with physics-based ASICs  (2507.10463 - Aifer et al., 14 Jul 2025) in Section 3, Subsection: Physical machine learning (Design Strategies)

First, device drift: PGLM coefficients can be updated with few-shot adaptation (e.g., tens of calibration circuits) to track slow drift; this is left to future work.

— Physics-Guided Linear Mapper for Quantum Error Mitigation  (2608.23109 - Chaudhari et al., 24 Aug 2026) in Section 6, Discussion