Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Develop effective learning algorithms for Physical Machine Learning (PML) and physical learning

Develop effective learning algorithms for physical machine learning that either (i) efficiently configure the controllable hardware parameters θ using a separate processor or (ii) realize on-chip physical learning in which the physical hardware itself performs the optimization of its own parameters.

Background

Physical machine learning (PML) seeks to co-design algorithms and hardware by directly optimizing physical hardware parameters so that the hardware’s intrinsic dynamics compute desired input–output mappings. In this paradigm, input vectors are encoded into a hardware system, which transforms them according to its physical evolution, and outputs are read from measured degrees of freedom; trainable parameters influence this transformation.

However, current PML demonstrations typically achieve only simple tasks due to two core difficulties: challenging optimization landscapes (e.g., barren plateaus) and a pronounced simulation-to-reality (sim2real) gap that undermines purely simulation-based training. Consequently, there is a need for learning algorithms that can either configure hardware parameters efficiently from a separate processor or, ideally, leverage the hardware itself to carry out parameter updates (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 Subsection "Physical machine learning" (Design Strategies)

In most of these systems the parameter update is still computed by an external digital processor, and performing the contrastive update within the array remains open.

— Two-State Dynamical Synapses as a Physical Substrate for Equilibrium Propagation  (2610.18058 - Liu et al., 13 Oct 2026) in Section 2, Related Work, subsection “Synaptic dynamics in energy-based and physical learning”