F-OAL: Forward-only Online Analytic Learning with Fast Training and Low Memory Footprint in Class Incremental Learning

Online Class Incremental Learning (OCIL) aims to train models incrementally, where data arrive in mini-batches, and previous data are not accessible. A major challenge in OCIL is Catastrophic Forgetting, i.e., the loss of previously learned knowledge. Among existing baselines, replay-based methods show competitive results but requires extra memory for storing exemplars, while exemplar-free (i.e., data need not be stored for replay in production) methods are resource-friendly but often lack accuracy. In this paper, we propose an exemplar-free approach--Forward-only Online Analytic Learning (F-OAL). Unlike traditional methods, F-OAL does not rely on back-propagation and is forward-only, significantly reducing memory usage and computational time. Cooperating with a pre-trained frozen encoder with Feature Fusion, F-OAL only needs to update a linear classifier by recursive least square. This approach simultaneously achieves high accuracy and low resource consumption. Extensive experiments on benchmark datasets demonstrate F-OAL's robust performance in OCIL scenarios. Code is available at https://github.com/liuyuchen-cz/F-OAL.

Paper

Similar papers

Peer review

Reviewer XB6P4/10 · confidence 4/52024-07-04

Summary

The author proposed a method called F-OAL for online class incremental learning, which does not rely on back-propagation and is forward-only, significantly reducing memory usage and computational time. In summarize, the contributions are as follows: 1) The paper presents theF-OAL, which is an exemplar-free method. 2) F-OAL can be updated in an mini-batch manner; 3) The methods are evaluated in several benchmarks;

Strengths

Strengths: The author proposed a method called F-OAL for online class incremental learning, which does not rely on back-propagation and is forward-only, significantly reducing memory usage and computational time.

Weaknesses

1) Some of the descriptions are unclear, such as Formula 4. The author may want to give a more vivid explanation; 2) The innovation is limited, See Limitations No 5; 3) The author may want to discuss more methods in evulation part;

Questions

See Limitations. The author may want to addresses the questions prpopsed in Limitations.

Rating

4

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

1. The formula No.4 seems to be wrong. Please check the formula carefully and give the proof process to ensure that the equation is valid. 2. Why does the formula No.4 show the optimal solution? For each mini-batch of data, the parameter W is the solution that makes the formula No.5 always equal to 0, in theory. But considering all batches of data, parameter W is too idealized and may even overfit the data of the current batch. 3. It is difficult for the parameter W calculated using a mini-batch of data to have an effect on other batches of data, especially when the distribution of data for different tasks is significantly different. Even if all the parameters W of the mini-batch are combined, I don’t think it can exceed the back-propagation based method in terms of effect, because this combination is linear. 4. Is the optimization process recursive on all batches of data? Why is the recursive method more efficient? Less computation? GPU parallel computing? Please explain the reason. 5. The innovation of this paper is insufficient. It seems that the main contribution is to calculate the parameter W by using the forward process and the least squares method. In fact, this method faces many disadvantages, such as overfitting. 6. Many models that appeared in the comparative experiments did not have annotated references, such as DVC in Table 1.

Reviewer UyXG5/10 · confidence 3/52024-07-10

Summary

This paper presents an analytic class incremental learning method that does not need backpropogation. The main idea is to use a pre-trained mode to extract features followed by random projection to higher dimensional space, and then use recursive least squares to update the linear regression weights. By doing so, the closed-form solution solves for all seen data thus guarantees no forgetting. The experiments in several class incremental image classification tasks show superior results than many continual learning baselines.

Strengths

- The empirical results of the method is very strong compared to other continual learning baselines.

Weaknesses

- Starting from section 3.2, it's better to give more precise definition for all notations, such as their dimensionality. - There are some confusing notations that can be improved. E.g., at line 121 you use $k$ to denote task, but at line 123 you use $k$ to denote batch, then in equation 9 you use change to $n$ to denote batch. - There are some writing issues like typos in the main, e.g. line 17, line 117, equation 4 (should be $\phi(X)^\top Y$?)

Questions

- Does the time recorded in Table 2 for F-OAL include the ViT feature extraction time?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations of the method are not discussed.

Reviewer nHHU8/10 · confidence 4/52024-07-17

Summary

The authors address the problem of online class incremental learning (OCIL), where new tasks arrive periodically in a data stream and the trainer seeks to learn these new tasks without catastrophic forgetting of past performance. The paper presents two modes of OCIL; replay-based methods and exemplar-free methods. Replay-based methods offer strong performance but require storing some amount of replay data from the stream to include in incremental training. Exemplar-free methods lift this limitation, but have not thus far achieved comparable performance to replay-based methods. The authors method (forward-only online analytic learning, or F-OAL) uses projections of blocks from a frozen, pre-trained ViT encoder as input to a trainable linear classifier, and then perform recursive least-squares updates on the linear classifier to avoid the catastrophic forgetting that backpropagation would cause in the same setting. The authors extensively compare their method by accuracy, training time, and memory usage to other recently-SOTA methods for OCIL, both replay-based and exemplar-free. The comparisons show that F-OAL significantly improves over existing exemplar-free methods and achieves comparable performance with replay-based methods, while being significantly more efficient in terms of both training time and memory usage. Lastly, the authors ablate their solution and find that keeping the ViT encoder frozen and using their analytically-learned classifier (instead of learning it through backpropogation) are both key ingredients in F-OAL's accuracy.

Strengths

The paper is generally quite strong. The F-OAL method is natural and intuitive; overall, it seems like a significant improvement to the Pareto frontier of accurate and efficient OCIL. The comparisons to past SOTA in both replay-based and exemplar-free methods are extensive and compelling. The authors' exposition is clear and logical, and the experiments are largely informative and useful for the reader. The F-OAL method seems pragmatic, and a natural baseline to which all future work in this area should be compared.

Weaknesses

This paper (and past papers in this vein) suggest that exemplar-free methods are "good for data privacy", but there is very little justification for this claim. I understand the basic premise as this: if replay-based methods force you to store some subset of the data stream, which is worse for users' data privacy than exemplar-free methods that don't require such storage. While it's worth noting as a design consideration/feature, I do not agree with this framing as relevant for "data privacy". There is no legitimate security model of privacy in the literature that would recognize this as "higher privacy". Once the model provider has seen/processed the data by running it through the model, any legitimate security model would view this as data that's been made public. Methods that enhance security (e.g. SMPC, HE, TEEs) and methods that provably reduce statistical privacy leakage (i.e. differential privacy) are orthogonal; these methods can be used interchangeably with both replay-based and exemplar-free methods! In a realistic setting, I can see how exemplar-free methods might help assuage compliance concerns or company-specific rules, but I know of no data privacy regulation or cryptographic threat model where this would be a relevant factor. I would suggest that the authors attempt to correct this misconception in the literature by simply stating the feature in terms of its utility: data need not be stored for replay in production. The reader can implicitly understand that this can have several benefits depending on the circumstances of the deployment. Otherwise, the only criticism I'd have for the paper is its ablation study. While it's a useful sanity check to see their result of ablating AC and FCC, the results with Frozen are obvious and unnecessary. The ViT-B model was developed for ImageNet-sized datasets, of course it will overfit CIFAR-100! I think there are more useful ablations that could be performed (more on that below).

Questions

It seems clear to me that the authors method of fusing the ViT blocks $B_i(.)$ with random linear projection is valuable, but also not entirely necessary. The goal of this approach appears two-fold; (1) capture information from different levels of abstraction in the representation that their analytic classifier uses, and (2) be able to control the dimensionality of that representation, which will surely need to be tune-able at training time for different datasets (e.g. to avoid overfitting). The ablation in Appendix B reassures the reader that (2) is necessary, but none of the ablations suggest that (1) is necessary. A simple ablation that could've helped would be to compare their block-averaging + smoothed projection approach with a simpler method that applies the smoothed projection to the last ViT block. I'd be curious to know why the authors chose this particular approach. The paper itself states that this feature fusion was implemented to "further enhance the representativeness of the features", but is there any work or experiment they can point to that suggests this? In any case, this seems to be the only weak point of the paper, and a clarification or improvement would be nice.

Rating

8

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The only concern I would have is that the paper (and its cited works) abuse the notion of "data privacy", which I have previously argued against in this review. Works that claim to "improve data privacy" without treatment of the staggering amount of literature that has gone into defining and proving what is and is not "private" in an information-theoretic, statistical, or engineering-focused sense are likely to muddy the waters for those fields. The paper's setting and solution are already worth publishing; in my opinion, the unsubstantiated data privacy claim is hurting more than it's helping.

Reviewer Pgch5/10 · confidence 4/52024-07-21

Summary

The paper introduces Foward-only Online Analytic Learning (F-OAL), an exemplar-free approach designed for Online Class Incremental Learning. The method addresses Catastrophic Forgetting by utilizing a pre-trained frozen encoder and a recursive least square updated linear classifier, which significantly reduces memory usage and computational time. The authors conducted extensive experiments to demonstrate the effectiveness of F-OAL on multiple benchmark datasets, showing its superior performance over existing exemplar-free methods and several replay-based methods.

Strengths

The F-OAL framework introduces a forward-only learning mechanism that avoids back-propagation, effectively reducing computational overhead and memory footprint. By not relying on exemplar storage, F-OAL maintains data privacy, a crucial requirement in many real-world applications where data sensitivity is a concern.

Weaknesses

While the paper is strong in many aspects, it lacks a detailed discussion on potential limitations of the proposed method, such as its dependence on the quality of the pre-trained encoder and the challenges that might arise in different data scenarios. Although the paper compares several baseline methods, including more recent and varied techniques could provide a more comprehensive evaluation of F-OAL’s relative performance. Some recent exemplar-free works could be easily generalized to OCIL setting and should be considered include: [1] Divide and not forget: Ensemble of selectively trained experts in continual learning, ICLR 2024 [2] R-dfcil: Relation-guided representation learning for data-free class incremental learning, ECCV 2022 [3] Self-sustaining representation expansion for non-exemplar class-incremental learning, CVPR 2022 [4] DiffClass: Diffusion-Based Class Incremental Learning, ECCV 2024

Questions

Please refer to the weaknesses.

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

Please refer to the weaknesses. I think providing a more comprehensive comparison with recent state-of-the-art works, a complexity analysis and an overhead comparison would help justify the contribution of this work.

Reviewer UyXG2024-08-11

I appreciate the authors' rebuttal. I still have minor concerns on the paper's presentation, as well as the limitation that it relies on a strong pretrained encoder. I will consider adjusting my score during the closed discussion. Thanks!

Authorsrebuttal2024-08-12

Thank you for the response

Thank you for taking the time to read our response! This means a lot to us! Please let us know if there is anything more needed from us!

Reviewer nHHU2024-08-12

Response to authors

Thank you for acknowledging and updating based on my remarks, these improvements are satisfactory in my view. However, I misunderstood the scoring procedure here. My score was contingent on those remarks being addressed, so it's unlikely to increase. I'm interested in the ongoing conversation with Reviewer UyXG as well, I hope that will be resolved before the closed discussion.

Authorsrebuttal2024-08-12

Thank you for the response

Thank you for taking the time to read our response! We are glad that the response addressed your concerns.

Reviewer Pgch2024-08-13

Thank you to the authors for their detailed response. I have also checked the feedback from other reviewers and will adjust my score accordingly.

Authorsrebuttal2024-08-14

Thank you for the response

Thank you for taking the time to read our respose. Could you kindly adjust our score if our response addresses your cocerns, before the discussion closes? Thanks!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC