How to Train Your Resistive Network: Generalized Equilibrium Propagation and Analytical Learning

The gist: The authors develop an algorithm to exactly calculate gradients using a graph theoretic and analytical framework for Kirchhoff’s laws, introducing Generalized Equilibrium Propagation (GEP) as a perturbative viewpoint that unifies Equilibrium Propagation (EP) and Coupled Learning (CL), showing how to train resistor networks without needing a replica or control over all edges<ref:2602.03546#pg12>.

How it works

The paper addresses the challenge of training analog computing systems while respecting locality constraints imposed by physical hardware, which is often energy-intensive<ref:2602.03546#pg2>. Two-phase learning rules, such as Equilibrium Propagation (EP) and Coupled Learning (CL), address this mismatch by comparing two nearby steady states of the same system under the same inputs in both a free phase and a nudged (weakly clamped) phase<ref:2602.03546#pg2>.

The core mathematical framework is rooted in linear response theory from nonequilibrium statistical mechanics, where relaxation dynamics are written as a gradient flow, which is then modified by adding a small training field to induce a nudged energy function<ref:2602.03546#pg12>. The EP identity states that the leading-order two-phase finite difference in parameter derivatives is proportional to the derivative of an objective function defined via the nudge, specifically showing that for standard EP, it recovers the gradient of a training objective with respect to parameters<ref:2602.03546#pg12>.

Two Learning Paradigms

The paper contrasts vanilla two-phase learning (EP-style) with projector-based (analytical) learning rules, which bypass nudging entirely by differentiating the closed-form linear response map of passive resistor networks<ref:2602.03546#pg12>.

1. Vanilla two-phase learning involves comparing a free steady state to a nudged steady state, where the nudge is an explicit linear energy term, leading to an update that is a “difference-of-squares” rule in resistance form<ref:2602.03546#pg12>.

2. The projector-based learning rule differentiates the closed-form map to obtain an analytical gradient with respect to edge resistances, which can be implemented physically using a small number of voltage- and current-mode experiments that realize the response operator ΩA/R<ref:2602.03546#pg12>.

Generalized Equilibrium Propagation (GEP)

GEP is introduced as a perturbative viewpoint that unifies EP and CL by considering the nudged energy as F(β, θ, y) = Eθ(y) + n(β, θ, y), where n satisfies n(0, θ, y) = 0 and n(β, θ, y) = O(β)<ref:2602.03546#pg12>. This framework allows for a direct comparison between two-phase estimators and the analytical circuit gradient<ref:2602.03546#pg12>.

The key result of GEP is that the leading-order two-phase finite difference in parameter derivatives is proportional to the derivative of a generalized objective function defined by h(theta, y0), where h(theta, y) = 1/k! d k n / dβ k (0, θ, y)<ref:2602.03546#pg12>.

Circuits as Tunable Input–Output Maps

The paper establishes a linear operator viewpoint where the circuit implements the map v = -ΩA/Rs s<ref:2602.03546#pg12>. This allows for explicit inference: a voltage-mode experiment implements s ↦→ v (left multiplication by-ΩA/R), and reciprocal/current-mode manipulations can access the adjoint action Ω T A/R<ref:2602.03546#pg12>.

The induced input–output map W R = -γ PoΩA/Rp i satisfies rank(W R) ≤ min(Ei, Eo, C)<ref:2602.03546#pg12>. This rank bound is crucial for understanding the expressivity of the network<ref:2602.03546#pg12>.

Performance and Bounds

Experiments on classification and regression tasks demonstrate that projector-based training is typically more stable than two-phase learning, while achieving comparable performance when the latter succeeds<ref:2602.03546#pg12>. Furthermore, the analytical gradient remains unbiased under additive zero-mean observation noise E[ε x] = 0, unlike the two-phase estimator which incurs both a finite-β approximation error and an O(β) noise-induced bias<ref:2602.03546#pg12>.

The analysis also provides bounds on the input–output map implemented by the circuit projector, showing that for an input pattern applied on the designated input edges, the resulting output voltages satisfy v(output) ≤ s(input)2/√Rmax/Rmin<ref:2602.03546#pg12>.

Conclusion

The work concludes that the R-weighted cycle-space projector ΩA/R directly defines an exact gradient for learning in passive networks, and that Equilibrium Propagation and Coupled Learning appear as approximate implementations of a similar projector-based learning rule<ref:2602.03546#pg12>. The paper suggests clear extensions to noise, dynamics, and nonlinear devices, and to co-design of graph structure, edge selection, and learning rules<ref:2602.03546#pg12>.

How it works

The authors show that the standard two-phase current-squared contrastive update converges in the β → 0 limit to the gradient of a distinct objective function<ref:2602.03546#pg12>. The analytical gradient does not incur the finite-β approximation error of the two-phase estimator and remains unbiased under additive zero-mean observation noise, with remaining uncertainty arising only from variance due to observation noise and finite-sample averaging<ref:2602.03546#pg12>.

The paper further demonstrates that while both vanilla two-phase learning and Ω learning perform well in the noiseless setting, Ω learning has a significant advantage in the noisy setting due to the statistical bias in the gradient estimates of vanilla two-phase learning<ref:2602.03546#pg12>.

How it works

The paper shows that for margin-violating samples, a subgradient is given by-v T T (I − ΩA/R) diag(i)<ref:2602.03546#pg12>. The projector-based scheme implements this update by applying an output-supported error vector ε and computing Δ = (I − Ω T A/R)ε, where the action of Ω T A/R is realized by a reciprocal/current-mode experiment<ref:2602.03546#pg12>.

The resulting resistance update is then local: Δr ∝ iF ⊙ Δ<ref:2602.03546#pg12>. This mirrors the analytical structure in (155) and removes finite-nudge bias at the estimator level<ref:2602.03546#pg12>.

How it works

The paper shows that for any vector u ∈ R E, we can set the source voltages to be s = r ⊙ u, then measure the voltages v = -ΩA/R s. We then perform a Hadamard multiplication by r-1, the element-wise reciprocal of r, obtaining-Ω T A/R u = r-1 ⊙ v<ref:2602.03546#pg12>. This can be done completely locally, adhering to physical constraints<ref:2602.03546#pg12>.

The paper further notes that because ΩA/R is an idempotent (oblique) projector, Ω squared A/R = ΩA/R, compositions of biasing/measurement steps simplify algebraically<ref:2602.03546#pg12>. This makes it possible to analyze complex multi-step physical protocols as products of ΩA/R and Ω T A/R, providing a direct bridge between circuit experiments and the learning rules developed in the main text<ref:2602.03546#pg12>.

How it works

The paper shows that for a given target yT ∈ R Eo, the loss L(theta) attains a global minimum on Θ, and min theta ∈Θ L(theta) = 0 iff yT ∈ f(Θ)<ref:2602.03546#pg12>. This confirms the existence of a global minimizer for the training objective<ref:2602.03546#pg12>.

The paper also shows that for a fixed target yT ∈ R Eo, L(theta) = 0 iff f(theta) = yT, which is possible iff yT ∈ f(Θ)<ref:2602.03546#pg12>. This confirms the existence of a global minimizer for the training objective<ref:2602.03546#pg12>.

1. textbfOmega-based Analytical Gradient Training for Resistor Networks (Algorithm 2): The system can compute an exact, physically interpretable gradient for training edge conductances without needing a replica or control over all edges by implementing the analytical update: The resistance update is then local: Δr ∝ iF ⊙ Δ.

2. textbfBeyond Finite-Nudge Bias in Learning Rules: Generalized Equilibrium Propagation (GEP): This framework unifies Equilibrium Propagation (EP) and Coupled Learning (CL) by assuming a generalized objective function of the form l(theta) = 1/k! ∇ k β n(0,theta,y), allowing for a comparison between two-phase estimators on a common footing for comparison with the exact resistor-network gradients.

3. textbfRobustness to Partial Control: Improved Performance Under Limited Actuation: The projector-based estimator exhibits lower variance and more consistent performance improvement as access to edges increases, whereas the contrastive estimator degrades more sharply under partial control, implying that the control over all resistive states (e.g. updating the resistive value based on the gradient) is not strictly necessary.

4. textbfHandling Noisy Data: Unbiased Analytical Gradient for Regression: The analytical gradient remains unbiased under additive zero-mean observation noise because the observed analytical gradient satisfies E[Aˆ(x) x] = A(x), unlike the two-phase estimator which suffers from a noise-induced bias.

5. textbf Enhanced Stability and Convergence: Superior Training Landscapes: Experiments show that while both methods achieve comparable accuracy, the projector-based training exhibits lower variance and more consistent performance improvement in noisy settings, suggesting it is typically more stable than two-phase learning.

Machine learning is a powerful method of extracting meaning from data; unfortunately, current digital hardware is extremely energy-intensive. There is interest in an alternative analog computing implementation that could match the performance of traditional machine learning while being significantly more energy-efficient. However, it remains unclear how to train such analog computing systems while adhering to locality constraints imposed by the physical (as opposed to digital) nature of these systems. Local learning algorithms such as Equilibrium Propagation and Coupled Learning have been proposed to address this issue. In this paper, we develop an algorithm to exactly calculate gradients using a graph theoretic and analytical framework for Kirchhoff's laws. We also introduce Generalized Equilibrium Propagation, a framework encompassing a broad class of Hebbian learning algorithms, including Coupled Learning and Equilibrium Propagation, and show how our algorithm compares. We demonstrate our algorithm using numerical simulations and show that we can train resistor networks without the need for a replica or readout over all resistors, only at the output layer. We also show that under the analytical gradient approach, it is possible to update only a subset of the resistance values without a strong degradation in performance.