Equilibrium Propagation for Non-Conservative Systems

This paper proposes a framework that extends Equilibrium Propagation to arbitrary non-conservative systems by modifying the learning-phase dynamics to account for non-reciprocal interactions, thereby enabling the exact computation of cost function gradients and achieving superior performance compared to previous methods.

The Gist

The Big Picture: Teaching a Machine Without a "Backward Pass"

Imagine you are trying to teach a robot to recognize a cat in a photo. In the standard way we do this today (called "Backpropagation"), the robot looks at the photo, makes a guess, realizes it's wrong, and then sends a "correction signal" all the way back through its brain, layer by layer, to fix its mistakes.

The problem is that this "backward pass" is very hard to build in real physical machines (like biological brains or silicon chips) because it requires sending information backward in time or across long distances instantly.

Equilibrium Propagation (EP) is a smarter, more physical way to learn. Instead of a backward pass, the robot just relaxes into a "calm state" (equilibrium). It tries two slightly different scenarios:

1. Free State: The robot looks at the picture and guesses naturally.
2. Nudged State: Someone gently pushes the robot's final guess toward the correct answer.

By comparing how the robot's brain changed between these two calm states, it can figure out exactly how to adjust its internal settings to get better next time. It's like learning by feeling the difference between "what I thought" and "what I was nudged to think."

The Problem: The "Symmetry" Rule

The original version of this learning method (EP) only worked for systems that follow a strict rule: Symmetry.

Think of a conservative system like a ball rolling on a smooth hill. If the ball rolls from point A to point B, the path it takes is determined by the shape of the hill. If you reverse the path, the physics are the same. In a computer brain, this means if Neuron A talks to Neuron B, Neuron B must talk back to Neuron A with the exact same strength.

However, many real-world systems (and modern AI models) are not like a smooth hill. They are like a river with a current or a one-way street.

• Non-Conservative Systems: Information flows one way (like in a feedforward network where data goes Input → Hidden → Output, but never backward).
• The Issue: The old EP method breaks in these systems. It tries to use the "hill" math on a "river," and the learning calculations become wrong. The robot learns the wrong lessons.

The Solution: Two New Methods

The authors propose two new ways to fix this, allowing the "Equilibrium Propagation" method to work on these one-way, non-symmetric systems.

1. Asymmetric EP (AsymEP): The "Local Fix"

Imagine you are trying to balance a scale, but someone keeps secretly adding weight to one side (the non-symmetric part). The old method just ignores this and tries to balance it anyway, which fails.

AsymEP adds a tiny, local "counter-weight" to the scale.

• How it works: During the "Nudged" phase (when the robot is being pushed toward the right answer), the algorithm adds a special correction term. This term is calculated based on exactly how "lopsided" or "non-symmetric" the connections are.
• The Analogy: It's like a cyclist riding a bike with a flat tire. The old method just tells them to pedal harder. AsymEP adds a small, local adjustment to the handlebars to compensate for the flat tire, allowing them to ride straight and learn correctly.
• Result: This allows the system to calculate the exact correct gradient (the right lesson) even when the connections are one-way.

2. Dyadic EP: The "Double-Brain" Approach

If AsymEP is a local fix, Dyadic EP is a bigger architectural change.

• The Analogy: Imagine you have a complex machine that only works if you have two identical copies of it running side-by-side. One copy represents the "forward" flow, and the other represents a "backward" flow.
• How it works: The algorithm doubles the number of variables in the system. It creates a new, larger "energy landscape" where the two copies interact. In this doubled space, the messy, one-way river of the original system transforms into a smooth, symmetrical hill again.
• The Result: Because the math now works on this "doubled" system, the learning is perfect. It's a bit like using a mirror to make a one-way street look like a two-way street so you can apply standard traffic rules.

What They Tested (The Experiments)

The authors didn't just do math; they tested these ideas on real image recognition tasks (like identifying handwritten digits or clothes).

1. Symmetric Start: They started with networks that were symmetrical (like the old EP). AsymEP learned faster and got better results than the old methods.
2. Forced Asymmetry: They forced the networks to be very "one-way" (highly asymmetric).
   • The old method (Vector Field) failed miserably, getting results no better than random guessing.
   • AsymEP kept working perfectly, even when the network was completely one-way.
3. Feedforward Networks: This is the big win. Modern AI (like the ones in your phone) are usually "feedforward" (strictly one-way). The old EP couldn't train these at all. AsymEP successfully trained these networks, proving it can handle the architecture used in most modern AI.
4. Deep Learning: They tested on a complex dataset (CIFAR-10) with a deep network. AsymEP and Dyadic EP performed almost exactly as well as the standard "Backpropagation" method, which is the gold standard.

Summary

• The Problem: The cool "Equilibrium Propagation" learning method only worked on symmetrical systems, but real AI and physical systems are often asymmetrical (one-way).
• The Fix: The authors created AsymEP (which adds a local correction to the learning rule) and Dyadic EP (which doubles the system size to make the math work).
• The Outcome: These new methods allow this physical, brain-friendly learning style to work on the same types of networks used in modern AI, achieving results just as good as the standard, difficult-to-implement methods.

In short, they figured out how to teach a physical machine using "relaxation" and "local nudges" even when the machine's internal wiring is strictly one-way.

Technical Summary

Technical Summary: Equilibrium Propagation for Non-Conservative Systems

1. Problem Statement

Standard neural network optimization relies on error backpropagation, which requires a distinct backward pass, nonlocal error signal transmission, and explicit gradient storage. These constraints are difficult to reconcile with biological plausibility and physical implementations (e.g., neuromorphic or analog hardware), which typically operate through local interactions and continuous relaxation.

Equilibrium Propagation (EP) offers a promising alternative by formulating learning as a contrast between two stationary states of a dynamical system: a "free" phase and a "nudged" phase. However, the original formulation of EP is restricted to conservative systems, where dynamics derive from an energy function, enforcing symmetric interactions (e.g., $J_{ij} = J_{ji}$). This limitation precludes the application of EP to a broad class of models characterized by non-conservative forces and non-reciprocal interactions, including:

• Modern feedforward architectures (dominant in AI).
• Biological circuits.
• Physical systems far from thermodynamic equilibrium (e.g., nonlinear optical systems, active matter, exciton-polariton condensates).

Previous attempts to generalize EP to non-conservative systems, such as the Vector Field (VF) algorithm, fail to compute the exact gradient of the cost function. They provide an unbiased gradient only in the conservative limit; as the antisymmetric part of the Jacobian increases, the gradient estimation error grows, potentially leading to optimization failure (e.g., maximizing cost rather than minimizing it).

2. Methodology

The authors propose two mathematically equivalent frameworks to extend EP to arbitrary non-conservative systems: Asymmetric EP (AsymEP) and Dyadic EP. Both methods retain the core EP principle of using stationary states for inference and learning but modify the dynamics to recover the exact gradient.

2.1 Asymmetric EP (AsymEP)

AsymEP preserves the original inference dynamics but introduces a local corrective term during the "nudged" phase.

• Mechanism: In the nudged phase, the system evolves under an augmented force field. This field includes the original force $F$, the standard nudging term $-\beta \frac{\partial C}{\partial x}$, and a new correction term proportional to the antisymmetric part of the Jacobian ($A_J$) at the free equilibrium:
  $$\frac{dx}{dt} = F(x, \theta) - \beta \frac{\partial C}{\partial x} - 2A_J(x_0, \theta)(x - x_0)$$
• Gradient Recovery: This correction effectively transposes the Jacobian in the learning rule, ensuring that the difference between the nudged and free stationary states yields the exact post-synaptic term required for the true gradient.
• Locality: The correction term is spatially local because $A_J$ vanishes for unconnected neurons, and the state difference $(x - x_0)$ is available at the synapse.

2.2 Dyadic EP

Dyadic EP is a variational approach that maps the non-conservative dynamics onto a conservative system by doubling the state space.

• Mechanism: The original $n$-variable system is mapped to a $2n$-variable system $(z, z')$ defined by an energy function $H(z, z', \theta)$ and a cost function $D(z, z')$. The energy function is constructed such that the original dynamics are recovered on the diagonal ($z=z'$), while the off-diagonal direction encodes the non-reciprocal forces.
  $$H(z, z', \theta) = -(z - z')^\top F\left(\frac{z + z'}{2}, \theta\right)$$
• Learning: The system evolves to a saddle point of the augmented energy $H_T = H + \beta D$. The difference $z_\beta - z'_\beta$ serves as the error signal.
• Relation to AsymEP: AsymEP can be viewed as the first-order projection of Dyadic EP onto the original $n$-dimensional space. Dyadic EP allows for parallel execution of positive and negative nudging phases but requires doubling the physical degrees of freedom.

3. Key Contributions

1. Exact Gradient Computation: The paper provides the first framework to compute the exact gradient of the cost function for arbitrary non-conservative dynamical systems using equilibrium propagation, overcoming the limitations of the Vector Field algorithm.
2. Two Generalizations: It introduces AsymEP (a direct modification of dynamics with a local correction) and Dyadic EP (a variational doubling of the state space), proving their equivalence in the limit of infinitesimal nudging.
3. Feedforward Capability: The methods enable the training of purely feedforward networks, a scenario where previous EP-based methods (like VF) fail because they cannot propagate error signals backward without explicit backward connections.
4. Theoretical Unification: The work demonstrates that the variational principle behind EP is universal and can be applied to non-reciprocal forces by extending the state space or modifying the dynamics, bridging the gap between energy-based models and general dynamical systems.

4. Experimental Results

The authors validate their framework on MNIST, Fashion-MNIST, and CIFAR-10 using continuous Hopfield networks and convolutional architectures.

• Symmetric Initialization: On MNIST with symmetric initialization, AsymEP achieves higher accuracy and learns faster than both standard EP and the Vector Field (VF) algorithm.
• Structural Asymmetry: When the network is constrained to have a high degree of structural asymmetry (where EP is inapplicable and VF degrades):
  • VF Performance: VF performance collapses as asymmetry increases, dropping to chance levels (e.g., ~10% accuracy on MNIST at high asymmetry).
  • AsymEP Performance: AsymEP maintains robust performance across all asymmetry levels, including completely antisymmetric connection matrices.
• Feedforward Architectures:
  • In a purely feedforward setting, VF effectively trains only the last layer (acting as an Extreme Learning Machine), resulting in poor performance (~64% on MNIST).
  • AsymEP successfully trains all layers, achieving ~92.7% accuracy on MNIST.
• Deep Networks (CIFAR-10): On a deep convolutional network trained on CIFAR-10, both AsymEP and Dyadic EP closely track the performance of standard Backpropagation (BP), achieving ~89.7% and ~90.7% accuracy respectively, compared to BP's 90.7%. In contrast, VF collapses to chance level.
• Stability: Experiments suggest that non-conservative dynamics trained with AsymEP can suppress oscillations and remain stable even under strong asymmetry and constrained input projections.

5. Significance and Claims

The authors claim that this work opens new avenues for learning in neuromorphic hardware, dissipative physical systems, and neural architectures where asymmetry is intrinsic rather than incidental.

• Physical Implementability: By removing the requirement for weight symmetry and explicit backward passes, the proposed algorithms are more compatible with physical substrates (e.g., memristors, optical systems, active matter) that naturally exhibit non-conservative dynamics.
• Biological Plausibility: The methods rely on local interactions and continuous relaxation, offering a more biologically plausible mechanism for credit assignment compared to backpropagation.
• Universality: The Dyadic EP formulation suggests that the variational principles of equilibrium propagation are universal, applicable to any network operating in a stationary state, regardless of whether the underlying forces are conservative or non-conservative.

The paper concludes that while AsymEP introduces a local corrective force that may require specific physical mechanisms for implementation, and Dyadic EP requires doubling the state space, both provide a rigorous theoretical and practical pathway to training non-conservative systems with exact gradients.