acceptodds
Under review as a conference paper at ICLR 2027

Controlling Symmetry Drift in Forward-Only Continual Learning

Abstract

Edge devices need to adapt in the field despite limited access to backpropagation, making node perturbation an appealing alternative based on local activity, neuron-level perturbations, and scalar loss feedback. Yet persistent perturbation noise can drive parameter growth over long data streams and undermine a learner's ability to adapt to later tasks. We show that, to leading order in a frozen-state analysis, non-gradient perturbation noise produces a mean-zero diffusion along loss-invariant rescaling symmetries, whereas gradient updates have no first-order component along these directions. We introduce gauge-damped node perturbation (), which periodically corrects symmetry-charge displacement through local reciprocal rescaling of each hidden neuron's incoming and outgoing weights. For positively homogeneous activations, each positive rescaling preserves the current function exactly, making the correction prediction-unbiased; the unclipped update contracts the targeted charge for an isolated layer pair, while the guarantee depends on the architecture. Across continual-learning streams, substantially controls charge displacement on long ReLU runs and can improve late-task accuracy; tanh experiments and simulated quantized writes further show that the benefit depends on the activation and write protocol.

Primary area:
optimization
Keywords:
Lifelong Learning, Forward propagation, Node perturbation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uhcxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates symmetry-charge drift in forward-only continual learning using node perturbation. The authors theoretically characterize how perturbation noise induces charge diffusion along loss-invariant rescaling symmetries, potentially affecting long-term learning. To mitigate this issue, they propose Gauge-Damped Node Perturbation (GD-NP), which periodically applies function-preserving weight rescaling to restore the symmetry charge toward its initial value.

Critical strengths

None

Critical weaknesses
  1. Insufficient connection between symmetry drift and loss of plasticity. The paper provides a theoretical characterization of symmetry-charge diffusion under node perturbation and proposes GD-NP to suppress it. However, it remains unclear to what extent this drift contributes to the observed loss of plasticity. In particular, exact gradient-based optimization can also suffer from plasticity loss despite preserving the symmetry charge under gradient flow. Therefore, restoring charge conservation alone does not establish improved plasticity, and the paper lacks sufficient evidence isolating the contribution of symmetry drift to long-term learning degradation.

  2. Questionable justification for restoring the initial symmetry charge. GD-NP explicitly drives the symmetry charge toward its initial value b0b_0. However, preserving the initial charge does not necessarily yield balanced parameter scales or favorable optimization conditioning, let alone optimal plasticity. Moreover, deviations from b0b_0 are not inherently detrimental and may even lead to better-balanced parameterizations. The paper does not sufficiently justify why restoring b0b_0 is preferable to directly controlling parameter imbalance or effective update scales.

Questions for authors
  1. Could the authors provide more direct evidence that symmetry-charge drift is a significant contributor to plasticity loss under node perturbation, rather than merely a correlated phenomenon? In particular, how much of the observed late-task degradation can be attributed to charge drift?

  2. Why does GD-NP restore the initial charge b0b_0 rather than targeting balanced parameter norms b=0b=0 or directly controlling relative update scales? Could the authors compare these alternative correction objectives in terms of late-task accuracy?

Reviewer_Z4s3OpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

Deploying learning algorithms for deep neural networks on edge devices is very challenging due to their tight memory and energy constraints. The authors propose to address this by exploiting node perturbation (a backpropagation-free, forward-only learning algorithm). Their contributions are as follows:

  • they provide a formal analysis of the effects of perturbation noise on the first-order diffusion of the symmetry charge, in comparison with SGD;
  • they demonstrate that charge noise cannot be cancelled by scalar reweighting in the case of linearized loss feedback;
  • they analyse how symmetry displacement results in larger weight norms and hinders learning, because the relative updates of different weights become imbalanced;
  • they introduce gauge-damped node perturbation (GD-NP), an algorithm that periodically corrects displacement while preserving the function represented by the network, along with a series of experiments to evaluate its effectiveness.
Critical strengths
  1. The paper addresses a topic relevant to the community, at the intersection of on-device learning, continual learning, alternatives to backpropagation and the understanding of learning dynamics. The authors identify a mechanism through which node perturbation causes parameter drift, and the formalism they introduce is sound within its stated assumptions.

  2. The authors provide a substantial appendix with full derivations, experimental protocols and complete result tables, and they are transparent about the limits of their claims. They also provide a well-documented codebase that regenerates the paper’s tables and figures from the logged results and gives the commands for the full training runs. This makes the claims easier to verify, although several contributions are presented mainly in the appendix rather than the main text.

Critical weaknesses
  1. The paper's clarity is lackluster. Many terms are introduced in the abstract and introduction without explanation and are defined only in the method section, or not at all. Examples include “model mismatch”, “frozen-state analysis”, “gauge”, “charge” and “orbit”. The authors never give the intuition behind the central concept of symmetry charge, nor explain how their four contributions relate to each other, so the paper is tedious to read and its main ideas are hard to extract. Several elements also suggest an unfinished draft: Figure 1 is never referenced in the main text, “Eq. equation” appears throughout, and Table A10 overflows the page. Moreover, contributions 2 and 3 are presented mostly in the appendix. As a result, the paper is very difficult to evaluate in its current form. A substantial rewrite of the abstract and introduction, giving intuitions before technical details, would drastically improve the paper's readability.

  2. The experimental support is limited (although acknowledged by the authors in the limitations). The experiments are small-scale:

  • small MLPs trained on data subsets;
  • mostly 20-task streams, with only one 100-task study;
  • permuted MNIST and grayscale, flattened CIFAR-10;
  • hardware effects only simulated.

The hyperparameters α and k are fixed without justification or sensitivity analysis. More importantly, the experiments do not show that symmetry drift is what limits learning. Displacement and accuracy are largely decoupled (weight decay leaves more drift than NP yet performs better), GD-NP alone does not outperform simple baselines such as weight decay or shrink-and-perturb, and it fails on tanh. As such, the practical value of GD-NP, and the relevance of the charge analysis to plasticity, remain unclear.

Questions for authors
  1. Appendix D introduces computational cost. Could the authors give the total memory and energy footprint of GD-NP compared with plain NP and with backpropagation?

  2. What does q(x) in 3.1 represent? Could the authors give its interpretation in the main text?

  3. The paper relies on several mathematical quantities: the charge bb, the displacement Dt\mathcal{D}_t, and the rates DBD_B and DbD_b. It does not explain what they correspond to in the network’s actual learning dynamics. Could the authors give an intuitive interpretation of each, e.g. what a large charge or a large displacement means for a neuron and for its ability to keep learning?

  4. The paper states that the leading-order model stands “if the state and noise statistics remain approximately constant”. Have the authors checked how well this holds during actual training?

  5. How should we understand that "E ∥Bt − B0∥ 2 F ≈ tDB"?

  6. Not all readers are familiar with continual learning. Could the authors define what a “task” is in each benchmark? Could they also define “period”, which seems to have two meanings: the correction period k of GD-NP, and the write or projection period of the quantized-write simulation in Section 7?

Reviewer_Uku8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents an approach to correcting a well known issue with the (zeroth-order) node perturbation (NP) method, namely the diffusion of network parameters in directions orthogonal to the gradient. The carry out a "charge-level analysis" and point out that positively homogeneous networks (e.g. MLPs without biases and with relu activations) have a gauge transformation (symmetry) such that the inputs to a node can be rescaled as long as the outputs are reciprocally rescaled and thus a network's function is unchanged. These directions are therefore flat directions which are always orthogonal to the true gradient, and yet NP reliably diffuses along these directions disrupting network training on the long term horizon. They propose a fix to this problem by tracking (from initialization) and applying a gauge-dampening (GD-NP) method to keep networks from drifting off in this space. They show that this effect is present in networks, that large deviations into function-preserving, null-gradient, directions on the parameter landscape can have catastrophic effects on learning, and that stability of training can be recovered. They also show some (modest) increases in final accuracy in continual learning tasks.

Critical strengths
  1. This paper has a very clear theoretical contribution. The core insight is clean, i.e. that gradients have no first-order component along exact symmetry directions, but NP’s does. The mathematical derivations are clear and make sense.

  2. The correction method proposed (gauge dampened node perturbation) is an exact and local correction compatible with perturbative training methods. It is exactly function-preserving for positively homogeneous networks, unlike weight decay. That makes it a natural fit for forward-only (inference) hardware, and the framework could be easily extended (to attention layers, batch norms etc).

  3. The paper reports results in a seemingly honest manner. Specifically I'm glad to see that it shows the results for tanh (a non positively homogeneous network) and the damping collapsing effects. Showing a failure mode of this method helps to make clear where (and for what) it is useful.

Critical weaknesses
  1. Though the method clearly shows that the gauge-dampening works and keeps the "balancedness displacement" low (Table 1), its effect on accuracy seems very limited. It would be good to have had a deeper inspection of whether the benefit of this gauge-dampening is solely for network stabilization over long training runs. If so it can be limited in it's contribution and application in that regime. The minor improvements over NP + decay seen in the results seem to indeed suggest that GD-NP does not have utility for increased model accuracy.

  2. The contributions of this paper are limited to some very small scale empirical results. It would have been good to have seen a much broader application of this method beyond the simple permuted-MNIST and CIFAR-10 experiments. Especially given that the literature by now contains many examples of NP like methods being applied to harder tasks than these. Again, this would be of interest in order to really understand how to see the impact of this gauge-dampening and it would be good to know if this is a minor aspect of the failure of NP in wider networks (which you appendix seems to suggest).

  3. The wider applicability of this method is unclear since it really only applies for positive homogeneous networks. It clearly fails for tanh models and so it is important to ask how far the impact of this method can be. One additional set of network components it could have been tested with (even in isolation as part of a larger network) could have been batch-norm layers, or attention layers.

Questions for authors

Since weight decay contracts your symmetry charge (b) toward zero, while the loss maintains the function (i.e. might push back on decay in the symmetry charge), decay should drive networks toward the some kind of balanced b rather than toward the initial b, correct? How does the actual evolution of b look for NP or NP + decay? And shouldn't GD-NP take this into account when being combined with decay?

Also, note that your abstract is being poorly displayed in your submission's abstract, there is a latex function "\ours{}" that is not resolving to "GD-NP".

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.