acceptodds
Withdrawn from ICLR 2027

Gated Target Propagation for Compositional Generalization in Continual Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual learning is typically framed as acquiring new knowledge without catastrophically forgetting previous tasks. However, a flexible continual learner should also be able to reuse and recombine previously acquired knowledge to rapidly solve novel task compositions. We introduce Gated Target Propagation (GaTaP), a continual learning algorithm in which task-specific gating variables–learned through a closed-form inner loop update–selectively suppress or enhance network modules. Network parameters are learned in a slower timescale outer loop, using the same local difference target propagation error signal as is used for adapting gating variables. We provide tractable experiments on class-incremental learning scenarios for both multilayer perceptron and convolutional network architectures. We show strong performance retention on previously learned tasks, as well as compositional generalization to unseen tasks, achieved through few-shot gain adaptation at inference. We analyze learned gating patterns and find that related tasks exhibit similar gating patterns, suggesting that inferred gates capture meaningful, reusable task structure. Overall, GaTaP provides a powerful framework for jointly ameliorating catastrophic forgetting and enabling few-shot compositional generalization in neural network models.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual learning, catastrophic forgetting, compositional generalization, target propagation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_zC2BOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

GaTaP combines shared network weights with task-dependent multiplicative gains. Gains are inferred from labeled support examples using sparse local regression, while shared weights are updated through probabilistic target propagation. Evaluation freezes the shared weights and re-infers gains for seen tasks and unseen combinations of classes.

Critical strengths

The separation between shared representation learning and support-based gain adaptation is clearly motivated, and the coordinate-wise sparse update is explicit. Evaluating unseen class combinations with frozen shared weights is a useful test of whether the representation supports rapid reuse. The support-size experiments, convolutional extension, and additional multi-class results go beyond a single MNIST demonstration.

Critical weaknesses

The evaluation differs materially from standard class-incremental learning. Algorithm 1 supplies a labeled support set for each evaluation task and re-infers its gains, including for previously learned tasks. This measures retention after supervised readaptation, rather than prediction among all seen classes without task-specific adaptation. Preserving original class labels does not remove this distinction. The task information, candidate output classes, and support-data access need to be specified consistently. The comparisons do not isolate the proposed learning rule. GaTaP receives a test-time adaptation mechanism, but it is unclear whether backpropagation and EWC receive an equivalent support-based adaptation opportunity. A frozen-feature classifier and a backpropagation-trained network with the same gains and support budget are important controls. Strong performance on new binary class combinations alone does not establish that modular composition, rather than ordinary transferable features, explains the gain. Equation (15) is a closed-form coordinate minimizer conditional on other gains and the inferred targets. It is not generally a single closed-form solution to the coupled multi-layer problem. The main text describes a single optimization step, whereas Appendix A.5 specifies ten gamma-inference steps at training and testing. This affects the efficiency claim and should be reconciled, including how targets are refreshed and convergence is assessed. The hyperparameter tables specify random seed 42, and the main results do not establish robustness across independent training seeds and task orders. The large reported gains need uncertainty estimates under the same adaptation protocol for every method.

Questions for authors

Which labels and candidate classes are available when evaluating an old task, and can the method predict jointly over all seen classes without a new labeled support set? How do matched gain-adaptation and frozen-feature baselines perform with the same support examples?

Reviewer_8s6COpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents Gated Target Propagation (GaTaP), this method helps neural network architecture to learn new tasks without forgetting what is already learned. The main idea is to use task-specific gates to choose which parts of the network are active, while keeping the earlier knowledge shared across tasks. The proposed gates can be updated efficiently using a closed-form solution, and the network weights are learned through local target propagation. The authors show that the proposed model (GaTaP) performs well on small scale MNIST, CIFAR-10, and CIFAR-100. GaTaP shows it is able to retain knowledge from older tasks while adapting to new combinations of tasks using very few examples only.

Critical strengths

In this work authors proposed to solve the catastrophic-forgetting problem by showing that the previous learned representation can be used to solve unseen task combinations. The paper shows that with the help of few-shot gate adaptation using the earlier frozen shared weights.

The paper proposes a task-dependent multiplicative gains mechanism which can be updated using the closed-form solution with the help of coordinate-wise optimization. This is very helpful to solve the problem without using the gradient based optimization and method can choose the relevant computational path without gradient descent.

This work leverages the kalman-style correlation to choose the local hidden target and use the layerwise objective to train the shared weight. The result on the small scale dataset shows the good generalization and retention for the image classification problem.

Critical weaknesses

The proposed approach builds on the many existing ideas, modular network routing are proposed in [a,b,c,d], task-dependent gating [e], and target propagation [f,g,h]. Also, the work ANML [e] already explores neuromodulated gating for selective plasticity in continual learning, and MNTDP [d] investigates the reuse and composition of learned modules. Kalman-style equation for the target updates is similar to the target-propagation methods [f,g]. I request the author to clarify its key novelty and show the advantages of combining these components.

The experiments are shown only on the MNIST, CIFAR-10, and CIFAR-100 dataset using relatively small MLP and CNN architectures without any standard architectures. This dataset and architecture are not sufficient to mimic the complexity of the real world. The author should include the result over the standard recent benchmarks for example Tiny-ImageNet, ImageNet-100 dataset with the ResNet or ViT architectures. The evaluation should be also for the longer sequence and multi-class settings.

GaTaP combines multiple components like sparse gating, local target inference, and bilevel optimization, but there is no ablation experiment to separate the individual components contributions. It is not unclear if the improvements come from target propagation or from task-specific gating. Which is the most important component? The component wise ablation is missing. I request the author please conduct the same to justify the each individual component.

There are no strong baseline in the paper, the author compared the result with the EWC paper which is published in 2017 (10 years older). Recently there are a lot of development in the continual learning domain and recent work shows much stronger results. Use the recent strong baseline to compare the results.

The claim and reported compositional generalization results do not fully establish in the paper it is not clear that the model has learned independently reusable concepts. A model with a strong shared representation learning and a simple added classifier over the top may give similar results.

References:

[a] Fernando et al., PathNet: Evolution Channels Gradient Descent in Super Neural Networks, GECCO, 2017.

[b] Mallya et al., PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning, CVPR, 2018.

[c] Serrà et al., Overcoming Catastrophic Forgetting with Hard Attention to the Task, ICML, 2018.

[d] Veniat et al., Efficient Continual Learning with Modular Networks and Task-Driven Priors, ICLR, 2021.

[e] Beaulieu et al., Learning to Continually Learn, ECAI, 2020.

[f] Lee et al., Difference Target Propagation, ECML PKDD, 2015.

[g] Meulemans et al., A Theoretical Framework for Target Propagation, NeurIPS, 2020.

[h] Ernoult et al., Towards Scaling Difference Target Propagation by Learning Backprop Targets, ICML, 2022.

[i] Buzzega et al., Dark Experience for General Continual Learning: a Strong, Simple Baseline, NeurIPS, 2020.

[j] van de Ven et al., Three Types of Incremental Learning, Nature Machine Intelligence, 2022.

Questions for authors

Please refer to the weakness section, paper need a significant experiment to show the claims raised in the work. It need realistic data and model with the component wise ablation to justify the contribution of the proposed component.

Also, there are no recent baselines included, please take a look to it.

Reviewer_2DmiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a new continual learning mechanism with task-specific gating variables. The authors report retention on MNIST and CIFAR-10 binary task sequences, accuracy on unseen class pairings, and held-out classes.

Critical strengths

I didn't find anything special about this paper.

Critical weaknesses

This paper has been rushed for submission without careful proofreading. I am highly unlikely to change my mind. I have 3 major concerns that I don't think the rebuttal can address.

1. Framing and positioning overclaim.

You are overclaiming in so many places; here are two examples:

a. The introduction says existing CL work is “typically evaluated only through the ability to remember previously learned tasks.” This mischaracterizes the field. Many methods explicitly target forward transfer, feature/representation learning, and plasticity. Basically, the GEm paper that you cited introduced FWT (forward transfer).

b. The claim to “formulate continual learning as a joint problem of retention and composition” mostly relabels the protocol. Recombining classes from different tasks into new binary pairs is a re-evaluation of standard split benchmarks, not a new problem formulation. It basically redefines class-incremental learning, not a problem you identified.

2. Experiments cannot be reproduced as you haven't included the code.

Datasets are too small to support the generality claimed. The main results use only MNIST and CIFAR-10, and CIFAR-100 appears only in the supplement. Standard class-incremental benchmarks such as CIFAR-100 with 10/20 splits as a main result, TinyImageNet, ImageNet-100/ImageNet-R and CORe50 are missing. Non-standard backbones. The MLP has two hidden layers (64, 128). The CNN is a custom “Simple ResNet” with kernel size 5, not the backbones used in the class-incremental literature, such as reduced ResNet-18 (GEM/A-GEM, DER), ResNet-32 (iCaRL) or ResNet-18 on ImageNet subsets. This makes direct comparison to published numbers impossible. Given recent CL work on pretrained models (e.g., ViT backbones), a pretrained transformer comparison would also show whether gain inference scales. Single seed (42), no error bars or variance.

No compute or wall-clock comparison, even though the Kalman-gain computation is acknowledged as expensive. No ablation of target propagation versus plain backprop with the same gain inference, so it is unclear which component drives the results.

3. Bibliography errors.

Paper cites:

JM Bernardo, MJ Bayarri, JO Berger, AP Dawid, D Heckerman, AFM Smith, M West, et al. The variational bayesian em algorithm for incomplete data: with application to scoring graphical model structures. Bayesian Statistics, 7(453-464):210, 2003. While all authors are fabricated, please see: https://mlg.eng.cam.ac.uk/zoubin/papers/valencia02.pdf

Also, the paper cites:

Eric Elmoznino, Thomas Jiralerspong, Yoshua Bengio, and Guillaume Lajoie. Towards a Formal Theory of Representational Compositionality. June 2025. URL https://openreview. net/forum?id=fXCfT7ErvL.

Sarthak Mittal, Yoshua Bengio, and Guillaume Lajoie. Is a Modular Architecture Enough? October 2022. URL https://openreview.net/forum?id=3-3XMModtrx.

They don't have a venue.

  1. The protocol is not class-incremental, and the comparison is unfair.

Class-incremental learning requires predicting among all classes seen so far without task identity at test time. GaTaP instead receives a labelled support set (64 examples per class) from the evaluation task and re-infers Γ before prediction. That is task-identity information plus test-time few-shot adaptation. The setting is closer to task-incremental or meta-learning.

Minimum fair baselines would include: Linear probe or few-shot fine-tuning of the head on the same support set with a frozen backbone. Task-incremental baselines with comparable information. The gating/mask methods discussed in Related Work: HAT, PackNet, SupSup-style supermasks, MNTDP, RPS-Net, none of which are compared. The statement that GaTaP “does not require explicit knowledge of task transitions” conflicts with requiring single-task epochs and a task-specific support set. Am I wrong?

Also, try to add more buffer-free SOTA methods and compare with them, such as:

WSN: Forget-free Continual Learning with Winning Subnetworks Haeyong Kang, ICML 2022

SNV: Shapley Neuron Values for Continual Learning: Which Neurons Matter Most? Vahedifar, ICML 2026.

  1. The “modular subnetwork” story is not what is evaluated. Tables 2 and 4 set the number of subnetworks to 1 for MNIST and CIFAR-10. The main results therefore come from a single network with per-unit (or per-spatial-location) gains, not from recombining parallel modules as Fig. 1 and §3.1 suggest. With spatial gating, Γ has C×H×W entries per layer. That makes it a high-capacity task-specific parameter fitted at test time, which further weakens the compositional interpretation.

  2. Sparsity is claimed but not shown. α is 1e-5 (MNIST, CIFAR-10) and 1e-7 (CIFAR-100). At these values the soft threshold is likely nearly inactive. The paper reports no sparsity statistics for Γ, so “sparse routing” is unsupported.

  3. Internal inconsistencies about the inner loop.

The Results section and A.3.1 say a single closed-form step / single iteration is used. Tables 2, 4 and 6 list “Gamma inference steps (train/test) 10/10.” Algorithm 1, line 9, computes h_L “followed by Γ” but never says which Γ is used in the forward pass that produces h_L before Γ has been inferred. K_L in Eq. 9/81 is undefined, since there is no layer above L. How y is represented (one-hot? MSE?) is also not stated.

  1. This weakness took so much time to identify each error, and it showed me that you didn't carefully proofread your paper. Mathematical errors and notation problems.

Laplace prior: p(Γ) ∝ exp(+α‖Γ‖₁) (Sec. 3.1) is not normalizable. It should be exp(−α‖Γ‖₁). Eq. 42 then uses the correct sign.

Posterior structure: Q(r, Γ | s, o) is per-sample, yet Γ is said to be constant across all samples of a task. The posterior over Γ should condition on the whole support set.

Noise inside losses: the random noise ε_i appears inside deterministic residuals and objectives (Eqs. 12, 13, 16, 69–72, 87, 93–94, 99–100). It should not be there. Eq. 45 also writes Σ₁ = ε₁I, conflating the noise variable with its variance.

Inconsistent gating operator: Eq. 2 uses Diag(Γ)W, while Eq. 66 uses Γ ⊙ W (vector ⊙ matrix, a dimension mismatch).

Dimensions: A.1 states W₁ ∈ R^{d₂} × R^{d₁}; it should be R^{d₁×d₀}. Sec. 3.1 says Γ_i^(k) ∈ R^{d_i}, while A.4.1 gives Γ ∈ R^{C_i} in the bullet list and R^{C×H×W} a few lines later.

Not truly bilevel: Eq. 18 writes Γ*(W), but Eq. 86 states that gradients are not propagated through Γ*. The procedure is alternating optimization, not bilevel.

Broken reference: “The full algorithm … is given in [1].”

Minor error:

Figure 1 takes roughly half a page to show a sum of row-gated matrices, which Eq. 1–2 already conveys. It could be a single-column schematic.

I believe weaknesses 1- 3 cannot be addressed during rebuttal. I can not verify the paper's results and numbers if you provide code in such a short time. The other weaknesses can be addressed. I am waiting for your response, and if I have proven I am wrong, I will likely increase my score.

Questions for authors

How do Backprop/EWC perform when given the same 64-shot support set at test time, for example via head fine-tuning or a linear probe on frozen features? Is the output head shared across tasks (10-way) or binary per task, and is it the same for all methods? Which Γ initializes the forward pass before Γ* is inferred? Is it one closed-form step or 10? What fraction of Γ entries are exactly zero after inference? What happens with n_i > 1 on MNIST/CIFAR-10? Do different subnetworks actually specialize?

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

Related papers

Discussion (0)

Sign in to comment.