Feedback-Regulated Residual RNN (FRE-RNN)

• The paper introduces FRE-RNN, which scales feedback through coefficients to reduce convergence time and computational cost while preserving trainability.
• It integrates residual connections to maintain gradient flow in deep networks and enables biologically plausible, local learning via equilibrium propagation.
• Empirical results demonstrate FRE-RNN achieves comparable accuracy to backpropagation while significantly reducing the number of iterations required for convergence.

Searching arXiv for the FRE-RNN paper and closely related equilibrium propagation work to ground the article in current literature. Feedback-regulated REsidual recurrent neural network (FRE-RNN) is a brain-inspired recurrent architecture introduced to make Equilibrium Propagation (EP) substantially more practical in large-scale networks by accelerating convergence, reducing iteration cost, and preserving trainability in deep settings (Liu et al., 5 Aug 2025). In this formulation, feedback pathways are explicitly regulated through coefficients $B_i$ so that the recurrent dynamics settle more rapidly, while residual connections are added to maintain information and gradient flow when feedback is weak. The architecture is used as the dynamical substrate for the two EP phases—a free phase and a weakly clamped phase—so that learning remains local and biologically plausible while mitigating the instability and computational expense that had limited earlier EP implementations (Liu et al., 5 Aug 2025).

1. Origin and problem setting

FRE-RNN was proposed in "Toward Practical Equilibrium Propagation: Brain-inspired Recurrent Neural Network with Feedback Regulation and Residual Connections" (Liu et al., 5 Aug 2025) in response to a central bottleneck in EP: existing implementations are described as slow and often unstable, with recurrent dynamics that may require dozens or even hundreds of iterations to reach a stable state. Because EP requires two equilibrium phases per sample, long settling times directly inflate computation and wall-clock training time. The paper identifies this as a practical obstacle both for large-scale neural networks and for hardware implementations (Liu et al., 5 Aug 2025).

The architectural intervention is targeted rather than generic. The paper argues that conventional EP dynamics are insufficiently controlled: if the recurrent system has too large a spectral radius, the state can oscillate or diverge, whereas if the spectral radius is too small, learning signals may vanish. In deep recurrent networks this creates a trade-off, because weak feedback improves convergence but worsens the vanishing gradient problem. FRE-RNN is presented as a structural remedy to this tension by regulating feedback rather than feedforward weights, and then compensating for weakened feedback through residual pathways (Liu et al., 5 Aug 2025).

A common misunderstanding is that faster EP convergence must be obtained by shrinking the entire recurrent or forward transformation. The paper explicitly rejects that design choice: it scales only the feedback weights $B_i$, with the stated purpose of lowering the effective dynamical gain of the recurrent loop while preserving forward signal propagation (Liu et al., 5 Aug 2025).

2. Core architectural design

FRE-RNN is a recurrent neural network in which feedback strength is explicitly scaled down, residual connections are introduced to preserve trainability under weak feedback, and the overall organization is framed as brain-inspired (Liu et al., 5 Aug 2025). In the layered formulation, the paper departs from the prototypical EP setting by separating the input and output layers from the recurrent hidden network, and the hidden state evolves in discrete time according to

$$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$

Here $p$ is the activation function, $W$ is the recurrent forward-weight matrix, and $b$ combines the feedforward input and error-nudging terms (Liu et al., 5 Aug 2025).

For the main algorithmic instance with two hidden layers, the free phase uses feedback clamped off, $B_f = 0$, whereas the nudged phase uses weak feedback, $B_f = B_{f1}$, with default $B_{f1} = 0.1$ (Liu et al., 5 Aug 2025). The paper gives the hidden-layer relations in the form

$h_1 = W_0 \cdot s_0 + B_1 \cdot B_1 \cdot s_2,$

$B_i$0

$B_i$1

with

$B_i$2

In this description, $B_i$3 is the target label, $B_i$4 is the prediction, and $B_i$5 is the prediction error. The stated functional role of feedback regulation is to scale the error-carrying influence into the recurrent network rather than weaken the forward drive itself (Liu et al., 5 Aug 2025).

Residual structure is introduced because weaker feedback, while beneficial for convergence, attenuates gradient flow in deep networks. The paper therefore adds cross-layer residual links that bypass neighboring layers. In a 10-hidden-layer symmetric model, three long-range bidirectional residual connections are added; in asymmetric networks, skip-layer connections are inserted between non-adjacent layers with 20% probability, yielding what the paper calls an arbitrary graph topology (AGT) (Liu et al., 5 Aug 2025).

3. Equilibrium Propagation formulation and local updates

Within FRE-RNN, learning is still performed in the EP framework. The network first reaches a free equilibrium, then a second equilibrium under weak output nudging, and the difference between the two steady states defines the local synaptic update (Liu et al., 5 Aug 2025). The paper states the update rule as

$B_i$6

and for the output layer,

$B_i$7

These rules are explicitly described as local and STDP-compatible (Liu et al., 5 Aug 2025).

The significance of this formulation is that FRE-RNN changes the dynamics without abandoning the core EP learning mechanism. The paper’s claim is not that EP is replaced by a new optimizer, but that the recurrent substrate is reorganized so that the two EP phases converge more quickly and more stably while still supporting local learning (Liu et al., 5 Aug 2025). This suggests that the main contribution lies in reconciling biological-plausibility constraints with a more tractable dynamical regime, rather than in redefining the EP objective.

The appendix further analyzes the weak-feedback or infinitesimal inference limit. For the nudged change in hidden state, the paper gives

$B_i$8

and for an earlier layer,

$B_i$9

This leads to the local update

$$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$0

With symmetric feedback, the derivation is described as recovering a BP-like gradient form up to a constant factor, which serves as the paper’s theoretical justification that weak feedback can preserve learning quality while improving convergence (Liu et al., 5 Aug 2025).

4. Spectral radius, Lyapunov analysis, and convergence behavior

The central dynamical claim of FRE-RNN is that reducing the spectral radius improves stability and convergence (Liu et al., 5 Aug 2025). The paper states that, in recurrent systems, the spectral radius is the largest eigenvalue magnitude of the weight matrix; if it is less than 1, the dynamics are stable and convergent, whereas if it is too large, the dynamics may become oscillatory or chaotic (Liu et al., 5 Aug 2025). FRE-RNN therefore regulates the recurrent loop by scaling feedback weights $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$1, which lowers the effective dynamical gain without degrading feedforward representational quality.

The appendix formalizes this through Lyapunov analysis. The Jacobian is written as

$$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$2

and the paper computes the maximum Lyapunov exponent (MLE) and finite-time maximum Lyapunov exponent (FTMLE) from this Jacobian (Liu et al., 5 Aug 2025). Smaller feedback scaling $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$3 is reported to correspond generally to lower (FT)MLE, indicating more stable dynamics and faster settling.

The practical implication is direct because EP training requires two equilibrium phases for every sample. The paper states that reducing feedback strength lowers both the number of iterations $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$4 needed for the free phase and the number of nudging steps $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$5 needed in the second phase, and that this can reduce total training time by orders of magnitude (Liu et al., 5 Aug 2025). A specific reported example is that, under small feedback scaling, a model with $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$6 and $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$7 can perform as well as one with $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$8 and $$s^{f}[t+1] = F(s^{f}[t], b) = p(W \cdot s^{f}[t] + b). \tag{2}$$9 (Liu et al., 5 Aug 2025). This does not imply that recurrence ceases to matter; rather, it indicates that the recurrent map becomes more contractive, so perturbations decay rather than amplify during phase settling.

5. Residual connectivity, arbitrary graph topology, and biological interpretation

Residual connectivity in FRE-RNN is motivated by a secondary consequence of weak feedback: the vanishing gradient problem in deep recurrent networks (Liu et al., 5 Aug 2025). The paper’s argument is architectural rather than algorithmic. Weak feedback is retained because it improves convergence, but residual and skip connections are added to ensure that signals do not attenuate excessively across many layers.

For a 10-hidden-layer network, the paper reports that weak feedback alone degrades performance, whereas adding residual connections restores substantial accuracy (Liu et al., 5 Aug 2025). It further states that the residual design can fully restore classification performance in deep models and that arbitrary residual links can surpass feedback alignment (Liu et al., 5 Aug 2025). The paper summarizes this interaction succinctly: feedback regulation makes EP fast, while residual connectivity makes deep EP trainable.

The topology is explicitly framed as brain-inspired. The authors draw analogies to cortical feedforward/feedback alternation, the neocortex’s extensive lateral and feedback loops, long-range recurrent and skip-layer projections, and a small-world-like recurrent organization (Liu et al., 5 Aug 2025). In that sense, the AGT construction is not presented merely as a graph-theoretic convenience; it is intended as a more cortex-like alternative to strictly layered networks.

A misconception worth separating from this claim is that any feedback-rich network is equivalent to FRE-RNN. A useful contrast is provided by "Neural Networks with Recurrent Generative Feedback" (Huang et al., 2020), which also uses recurrent top-down feedback and iterative refinement, but does so through a deconvolutional generative model, explicit latent variables $p$0, MAP inference, and self-consistency constraints (Huang et al., 2020). FRE-RNN, by contrast, is described in terms of recurrent residual updates, feedback regulation, and EP-compatible local learning rather than an explicitly Bayesian generative posterior (Liu et al., 5 Aug 2025).

6. Empirical performance and practical significance

The experimental evaluation in (Liu et al., 5 Aug 2025) covers MNIST and Fashion-MNIST (FMNIST), layered RNNs with 2, 3, 5, and 10 hidden layers, convolutional RNN embeddings, symmetric and asymmetric recurrent connectivity, and arbitrary graph topologies with residual links. The strongest claims reported are that FRE-RNN improves EP training speed by at least one order of magnitude over prior P-EP for both layered and convolutional architectures, that accuracy is comparable to BP on benchmark tasks, that lower feedback scaling substantially improves convergence speed, and that residual links recover performance lost due to weak feedback in deep models (Liu et al., 5 Aug 2025).

The quantitative results reported in the paper include the following.

Setting	Accuracy	Time
2-hidden-layer, Ours (tanh, Adam)	training $p$1, test $p$2	1:16
2-hidden-layer, P-EP (sigmoid-s)	test $p$3	1:56
3-hidden-layer, Ours (tanh, Adam)	test $p$4	2:11
3-hidden-layer, BP (tanh, Adam)	test $p$5	0:24
Convolutional MNIST, Ours (hard-sigmoid)	test $p$6	12:28
Convolutional MNIST, P-EP	test $p$7	8:58
Convolutional MNIST, BP	test $p$8	1:01

These results are interpreted in the paper primarily as evidence of improvement relative to prior EP rather than relative to direct BP (Liu et al., 5 Aug 2025). For the 3-hidden-layer case, the reported test accuracy of FRE-RNN matches BP at $p$9, but BP remains faster in wall-clock time. The significance assigned by the paper lies instead in obtaining BP-level benchmark performance while retaining EP’s local learning structure and biological plausibility.

In deeper networks, the paper reports that plain asymmetric EP models underperform, whereas residual or AGT connections improve performance markedly (Liu et al., 5 Aug 2025). For the 10-hidden-layer case, it reports that residual connections bring about 5% absolute improvement on MNIST and 7% on FMNIST, and that the 10-hidden-layer residual symmetric model reaches around 97.5% test accuracy on MNIST (Liu et al., 5 Aug 2025). This suggests that the architectural combination, rather than feedback regulation alone, is necessary for scaling EP beyond shallow recurrent systems.

From the standpoint of application, the paper emphasizes two consequences. First, the reduction in required iterations makes EP substantially more usable in networks of practical size. Second, because the learning rule remains local and the dynamics are more naturally convergent, the techniques are presented as guidance for implementing in-situ learning in physical neural networks and for brain-inspired computing hardware (Liu et al., 5 Aug 2025). The broader implication is not that FRE-RNN eliminates the performance gap to BP in all respects, but that it narrows the practicality gap that had previously limited EP’s deployment.