acceptodds
Withdrawn from ICLR 2027

Dynamic Concept Pool with Instance-Adaptive Gating for Interpretable Class-Incremental Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual learning (CL) aims to acquire new knowledge from evolving tasks while preserving previously learned knowledge. Interpretable CL has attracted increasing attention for its potential to achieve strong continual learning performance while providing more transparent decision processes. However, existing methods often construct semantic concepts only once per incremental stage, limiting their adaptability to evolving class relationships, while final predictions still rely on trainable classification modules that may accumulate optimization errors over time. To address these limitations, we propose Dynamic Concept Pool with Instance-Adaptive Gating (DyCoG), an interpretable continual learning framework that grounds both knowledge accumulation and classification in explicit semantic concepts. Specifically, the Dynamic Concept Pool (DCP) progressively refines a human-understandable class-level concept space through classification feedback as new classes are introduced, enabling adaptive and interpretable knowledge accumulation. The Instance-Adaptive Concept Gate (IACG) further selects an instance-specific subset of concepts for each input, making the final decision directly traceable to instance-relevant semantic evidence. The resulting concept representations are directly used for statistical classification without relying on trainable classification modules for final prediction. Experiments on four benchmark datasets demonstrate that DyCoG delivers strong continual learning performance with competitive computational cost. More notably, as task partitions become finer, DyCoG exhibits an overall performance improvement rather than degradation, in contrast to the declining or saturating trends observed in most compared methods.

Primary area:
interpretability and explainable AI
Keywords:
continual learning, class-incremental learning, interpretable machine learning, concept-based learning, dynamic concept pool, instance-adaptive concept gating

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

DyCoG is an interpretable model for class-incremental learning. It preserves samples from previous phases to use for training in future phases. It uses an LLM to add new concepts to the concept bank and refine them. It also employs a gating mechanism to find relevant concepts for each image among all available concepts. It avoids using a gradient-based classifier and instead uses an RLDA classifier based on image-concept similarities.

Critical strengths

The paper addresses two fundamental problems:

  1. Interpretability: The ability to provide explanations for the model's classification decisions.
  2. Class-incremental learning: The ability to distinguish new classes while preserving classification performance on previously learned classes.
Critical weaknesses

Unfair comparison in Table1

DyCoG uses the training data from all past classes, while most of the compared methods are exemplar-free, such as CI-CBM, CLG-CBM, etc. These methods do not preserve any samples from previous classes/stages. It is obvious that having access to previous classes makes it much easier to preserve their classification performance. The main challenge in CIL is to learn to distinguish new classes with previous classes without access to the previous samples. This issue also affects the comparisons in Tables 2 and 3, and Figure 5 as well and leads to a false conclusion that DyCoG benefits from more incremental stages while other methods do not. The reason for this might not be the DyCoG architecture and methodology but could instead be due to DyCoG using all previous samples.

Questions for authors
  1. Does your method support the case where you do not retain any samples from previous classes? If yes, those results should be compared with other exemplar-free methods, and you can have another table to compare with methods that preserve a pool of previous samples. This will make the comparison fair.
  2. What is the difference between applying your method incrementally and training an interpretable classifier using all seen samples?
  3. While you mentioned that your method avoids gradient-based classifier training, your gating network is trained using gradient descent at each stage. Can you clarify this?
Reviewer_XXznOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

the paper introduces DyCoG, interpretable CIL method. It replaces trainable classification head with generated by LLM concept pool scored by CLIP and LDA is classifier method. QWEN generates concepts, LDA says whether they should be kept or not with hte use of a gate to select them. The method keeps all historical data, thus achieving best performance.

Critical strengths

Good performance and simplicity in terms of solution.

Critical weaknesses

The method is comparing itself mostly against exemplar-free methods (such as ICICLE) while keeping 100% HISTORICAL DATA. That is completely unfair for those methods, and it is not stated at all.

Then the interpretability part, it is also not holding. One image can use even up to 250 concepts (appendix A7)! who will analyze such enormous explanation. Claim about concept or interpretability if false.

Moreover, the gate has no sparsity when learning, it only learns ones and uses all that help accuracy. Also, on CUB the gate (one of the bolded highlighted contributions by authors) does nothing, it is inactive!

There is no judgement of concepts by humans nor by methods that are suitable to do so, they are just analyzed by LLM...

Baselines results are weird. L2P gets 0.52%, FlyPrompt 4.93%, and HippoTune 4.08% on CIFAR100 at T=100. These methods don’t collapse to chance with a pretrained ViT!

Questions for authors

How does you model work in exemplar free CIL?

Do random-concept and random-embedding baseline.

provide a plot on accuracy vs sparsity

Metrics on interpretability? Any user study?

Why baselines are so low?

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

Paper proposes a method for interpretable continual learning. The method has two phases. Phase I: Class-level concept modeling, in which a set of concepts is generated and managed. Phase II: Instance-level concept selection which learns a gating network to select the relevant concepts for each instance. The method uses an LLM to generate concepts, and then selects a set of them, after a conceptual cleaning phase. The method seems to have access to 100% of historical data (Table 2) which is in conflict with standard continual learning protocol. The results are in Table 1 are very good (I would think close to joint training).

Critical strengths
  • I would highlight the excellent results which is obtains in Table 1 (but there are some important question marks).
Critical weaknesses

Critical

  • The method seem to be requiring all historical data in Table 1 (see Table 2), this is in conflict with the standard setup for CL and should be clearly described in introduction, method and experiments.
  • The method is not well presented. From Figure 2 it seems that only DCP is being trained (indicated by a flame), however, I understand that there is no network involved. I think that there is a network involved in training GATE in stage 2, which is identified as G_t in the algorithm (line 13) in the appendix but never defined in the main paper. The design of G_t I could not find (MLP?).
Questions for authors
  • Please discuss your setting. Are you using full historical data access ? Also the method section seems to suggest this 'New candidates are merged with retained concepts and evaluated by Conceptual Cleaning using training images from all observed classes.' This something which is in conflict with standard CL settings. I am aware that some methods in CL study full memory, but this should be really clearly mentioned.
  • Could you also provide results for joint training of the method ?
  • I would prefer the results in the main paper also to include Last task average (or if one, only last task average). Especially in Table 2. I would be interesting to see if it goes down at all with increasing T (please comment this).
  • please explain why 'concepts are initialized only once' is not good. It would be good to be more precise in your motivation, because this argument is not directly understandable.

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

Related papers

Discussion (0)

Sign in to comment.