equilibrium-propagation

A post-connectionist blueprint for persistent, autopoietic AGI. This framework redefines intelligence as dynamical self-stabilization, moving from "AI as a tool" to "AI as a persistent entity" through recursive self-modeling, active inference, and non-equilibrium thermodynamics.

Empirical research toward persistent digital identity — 17 experiments, EqProp-reservoir synthesis, behavioral momentum. Toward Vivy: Fluorite Eye's Song.

Exploratory companion to arXiv:2601.18710: a quantum Ising-model classifier vs. Equilibrium Propagation on AML blood-cell images, with circuit visualizations.

Aspirational floating-gate arm of memristive-crossbar (02): can in-fabric learning run on a real analog floating-gate array (tunneling/injection writes)? Verdict-gated, H5-first, provisional.

A detailed-balanced chemical reaction network realized as a gradient flow, where equilibrium propagation is rigorous — the chemistry computes its own weight update (verified: local update == true gradient). Audits 'who computes the update?' across a co-location spectrum vs an offline ceiling. Part of physical-learning-substrates.

Decentralized, physics-driven coupled learning in resistor/flow networks - does the substrate's own physics compute the weight update?

Hardware-transfer arm of coupled-learning-networks (01): toward a 3D-printed flow network that learns by directed aging. Verdict-gated, H5-first.

The meta-project / commons of the physical-learning-substrates portfolio: records, adapts, and synergizes the common lessons of projects 01-05 and the self-assembly parent line. A knowledge commons + a substrate-agnostic code commons + a harvest tool, joined by a human-triggered synergy loop.

Hardware-transfer arm of memristive-crossbar (02): does in-fabric learning survive a real floating-gate FPAA? Verdict-gated, H5-first.

Equilibrium Propagation in PyTorch: train a net without backprop via two-phase energy relaxation. Tutorial + MNIST; companion to arXiv:2601.18710.