Change, memory, adaptation, and learning

Push a spring and it changes shape. Alter a memory device and it may retain a record. Train a model and its later performance may improve. These examples illustrate different claims. Hill’s idea becomes stronger when their differences remain explicit: response concerns dependence on current conditions; memory concerns accessible dependence on earlier conditions; learning concerns a retained change that improves a specified task through experience. Biological adaptation also involves mechanisms such as inheritance and selection, which need their own description. These are working distinctions for this research program, not a claim that every discipline uses identical terminology.

Performance must be defined independently

A learning demonstration needs an outcome that was chosen before seeing the result. Otherwise almost any state change can be declared improvement after the fact. We propose specifying a loss function, a training history, and an evaluation distribution. Physical-network research supplies concrete examples of systems that alter parameters to perform assigned tasks. Such examples establish that learning can be physically embodied in particular architectures. They do not imply that every object with history dependence learns or that an improvement on one task transfers to every other task.

Retention and generalization are different tests

A system may perform better immediately after training because of a temporary state, then return to its earlier behavior. It may also memorize the training examples without handling new inputs. We propose reporting immediate performance, delayed performance, and held-out performance separately. Training and evaluation should not share hidden labels or repeated samples that leak the answer. This creates a more useful comparison than a single success percentage, and it makes clear whether the reported effect is transient response, retained specialization, or broader generalization.

The minimum claim ladder

For each candidate system, first show a repeatable response. Then test whether different histories remain distinguishable under matched current conditions. Next test whether retaining that distinction changes a later outcome. Finally evaluate whether the change improves a prespecified task beyond appropriate controls. A result can stop at any level and still be informative. We propose publishing the achieved level rather than skipping directly to universal learning. The larger Hill hypothesis would need a mechanism connecting such local findings across scales.

The mathematical lens

Turn it into a question.

Follow the original work.

These sources support the mechanisms and limits discussed here. They do not endorse Hill’s framework. Notes distinguish research proposals, reviews, and observations.

1. Supervised Learning in Physical Networks: From Machine Learning to Learning MachinesMenachem Stern; Daniel Hexner; Jason W. Rocks; Andrea J. Liu · 2021 · Physical Review X 11, 021045Read original source ↗
2. Demonstration of Decentralized Physics-Driven LearningSam Dillavou; Menachem Stern; Andrea J. Liu; Douglas J. Durian · 2022 · Physical Review Applied 18, 014040Read original source ↗
3. Semantic information, autonomous agency and non-equilibrium statistical physicsArtemy Kolchinsky; David H. Wolpert · 2018 · Interface Focus 8, 20180041Read original source ↗