acceptodds
Withdrawn from ICLR 2027

Beyond the Observed: Knowledge-Informed Controllable Augmentation for Few-Shot Class-Incremental Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-Shot Class-Incremental Learning (FSCIL) addresses the practical need to continually acquire novel classes from limited labeled data while retaining previously learned knowledge. While previous studies primarily focus on model-side strategies for improving adaptation and knowledge retention, the role of data-side augmentation remains relatively underexplored. We argue that a key limitation of few-shot scenarios lies in the severely limited intra-class variations captured by sparse observations, and that external knowledge can provide useful priors for exploring plausible variations beyond the observed samples. Based on this insight, we propose Knowledge-Informed Controllable Augmentation (KICA) that derives plausible variation instructions from external class knowledge and realizes them through image-conditioned diffusion-based editing. By combining the visual evidence of few-shot observations with external variation knowledge, KICA preserves class-specific characteristics while enriching intra-class variations, achieving broader coverage of real intra-class variations as evidenced by distributional analyses. Extensive experiments across multiple FSCIL benchmarks and network architectures demonstrate consistent performance improvements, which validate the effectiveness of knowledge-informed augmentation and motivate a datacentric direction through expanded intra-class variation coverage.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Few-Shot Class-Incremental Learning, Generative Data Augmentation, Knowledge-Guided Learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_wFLAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Knowledge-Informed Controllable Augmentation (KICA) for few-shot class-incremental learning, inspired by the idea of augmenting samples for each novel class. Specifically, it first class-specific attribute sets from WordNet, Visual Genome, and an LLM, and then translates this information into structured instructions for background, attribute, and viewpoint editing. A pretrained image editor generates additional examples from the few-shot support images, which are subsequently used by existing FSCIL methods without modifying their architectures or objectives. Experimental results show that KICA can further improve their FSCIL performance.

Critical strengths
  1. The idea of generating novel samples for each class is reasonable.

  2. KICA exhibits consistent improvements over the original downstream methods. Table 1 reports higher final-session accuracy in all nine method–dataset combinations. Some improvements over plain image editing are substantial, particularly for BiMC on CIFAR-100 and CUB, suggesting that the proposed guidance can make synthetic augmentation more useful.

  3. The experiments are extensive and sufficient to show the efficacy of KICA.

Critical weaknesses
  1. The overall pipeline—constructing textual descriptions, converting them into editing instructions, and augmenting real images with a pretrained generator—seems incremental. For example, ALIA already uses vision and language models to construct descriptions and perform language-guided image editing for data augmentation [1]. The potentially distinctive contribution may be the integration of external class knowledge in KICA. [1] Dunlap, Lisa, et al. "Diversify your vision datasets with automatic diffusion-based augmentation." Advances in neural information processing systems 36 (2023): 79024-79034.

  2. Section 3.2 aggregates attributes at the class level, but an attribute associated with a class is not necessarily a valid replacement for every instance. Valid variations may depend on other attributes or biological factors, especially for fine-grained bird categories. A simple union of attribute values does not capture these dependencies, and textual preservation instructions do not guarantee that the image editor respects them. This matters because all generated images retain their original labels and are used without filtering.

  3. The paper states that results are averaged over repeated runs with paired seeds, but does not report the number of runs or uncertainty estimates. This is consequential because several gains over plain image editing are relatively small—for example, CLOSER’s final-accuracy improvements are approximately 0.7 percentage points across the three datasets, while BiMC improves by 0.2 points on miniImageNet. Reporting variability across support-set sampling and generation seeds would clarify their robustness. Table 4 also reports substantial preprocessing costs: approximately 20 hours for 800 images and 66.7 hours for 2,000 images on an A800 GPU, with roughly 76 GB of GPU memory usage. Although inference remains unchanged, these costs matter when new classes arrive incrementally. A generation-budget ablation would help establish how much of the improvement can be retained at a lower cost.

Questions for authors
  1. How much improvement remains when the plain editing baseline uses exactly the same operation-specific templates and preservation constraints as KICA, but without the external attribute base?
  2. What are the individual contributions of WordNet, Visual Genome, and LLM-derived attributes? Does the combined system outperform an LLM-only variant?
  3. What fraction of generated images preserve the original fine-grained label, particularly for attribute edits on CUB?
  4. How do the gains divide between novel-class recognition and retention of previously learned classes?
Reviewer_5bxNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the problem of few-shot class-incremental learning (FSCIL) from the data augmentation perspective. Specifically, it argues that FSCIL settings mainly fall short on the intra-class diversity, which could be mitigated by the powerful LLM-based image editing proposed by this paper. Extensive experiments on multiple benchmarks have demonstrated that it successfully augments several baselines with considerable performance gains and outperforms traditional augmentation approaches.

Critical strengths
  1. This paper is well-organized and easy to follow.
  2. Extensive experiments have demonstrated the effectiveness of the proposed method.
Critical weaknesses
  1. Using powerful LLMs to augment small FSCIL models forms an unfair comparison to general FSCIL methods and a less novel paradigm; at least, it fails to reach the standard of ICLR.

  2. I believe such a paradigm only works well on the well-defined categories in the current datasets. In other words, this method only defers the upper bound to LLM.

  3. I wonder how this paper controls the quality of the generated images.

  4. This paper should highlight in which scenario such an LLM-augmented FSCIL method is superior to directly using an LLM or a smaller distilled VLM. (e.g., less cost during the inference stage).

Questions for authors

Please refer to the Critical Weaknesses.

Reviewer_J8eQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes KICA, a data-level augmentation framework for FSCIL. It argues that few-shot novel classes cover only a limited region of the intra-class variation space, so conventional augmentation cannot recover plausible missing variations.

Critical strengths
  1. The method is modular and plug-in. KICA modifies only the support set and leaves downstream architectures, objectives, and inference untouched.

  2. The motivation is concrete and visually grounded. Fig. 1 makes the "limited coverage of intra-class variation" problem explicit.

Critical weaknesses
  1. The motivation's central premise is asserted more than measured. The paper claims that few-shot observations cover only a limited portion of intra-class variation, but the only direct evidence is Fig. 1, a schematic. The distributional analysis (Table 2) measures coverage of augmented sets, not the gap between observed support and true class distribution. A stronger motivation would quantify, per class, how far the 5-shot support lies from the full training distribution and show that this gap predicts where KICA helps.

  2. Efficiency cost is substantial and under-discussed. Table 4 reports ~20 hours for 800 images (CIFAR-100/miniImageNet) and ~66.7 hours for 2,000 CUB images, at ~90–120 s/image. The paper calls this "offline preprocessing" and notes no inference overhead, but for FSCIL the practical appeal is precisely low incremental cost. A 20–67 hour preprocessing stage per dataset is a significant barrier, and the paper provides no comparison to the wall-clock cost of the baselines it outperforms.

  3. Table 1 reports only A_last and A_avg. These can mask non-monotonic session behavior and cannot reveal whether KICA's gain is concentrated in early or late incremental sessions. More importantly, neither metric separates base-class from novel-class accuracy, so it is impossible to tell whether KICA improves novel-class learning without increasing forgetting of base classes.

Questions for authors
  1. Can you quantify the "limited coverage" premise directly, e.g., per-class distance between 5-shot support and the full class distribution in a fixed feature space, and show that this gap correlates with KICA's per-class gain?

  2. Please add a baseline that uses the same attribute knowledge as text-only conditioning (no image editing), to separate the contribution of the attribute base from the contribution of instruction-based editing.

  3. Please provide a quantitative or human evaluation of edit fidelity, e.g., the fraction of edited images whose predicted class (by a fixed classifier) matches the original, to support the "class-consistent" claim.

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

The paper studies few-shot class-incremental learning (FSCIL). It addresses the limited number of training samples for new classes and the lack of intra-class variation in these samples. The authors propose KICA, a knowledge-informed controllable data augmentation framework. KICA combines WordNet, Visual Genome, and a pretrained large language model (LLM) to construct a class-attribute bank. It then uses structured positive and negative prompts to guide Qwen-Image-Edit. The model performs background replacement, attribute editing, and viewpoint adjustment. It preserves the architecture of the downstream FSCIL model and expands the support set for each new class. The authors evaluate KICA on CIFAR-100, miniImageNet, and CUB. They apply it to three FSCIL methods: CLOSER, LGSP, and BiMC. They compare KICA with CutMix, RandAugment, text-to-image generation, and conventional image editing. KICA achieves the best results in 17 of the 18 reported evaluation settings. The authors also analyze the generated data with Coverage, adjusted IRS, and t-SNE. They report the offline computational cost of image generation. The paper addresses a relevant problem. Its data-centric approach also has practical value. However, KICA mainly combines existing knowledge resources, LLM-based prompt generation, and a pretrained image-editing model. The current experiments do not fully establish the independent contribution of knowledge guidance. The paper also lacks direct validation of the class labels assigned to generated images.

Critical strengths

1.The problem is practically relevant. The paper addresses insufficient intra-class variation in few-shot learning. Its data augmentation approach complements model-based methods. 2.The framework is modular. KICA preserves the main architecture of downstream FSCIL methods. The authors evaluate it with methods based on CNN, ViT, and CLIP. 3.The experiments cover several settings. The authors evaluate three standard datasets and compare KICA with conventional and generative augmentation strategies. The comparisons use the same number of additional images. 4.The paper analyzes data distribution and computational cost. In addition to classification accuracy, the authors report Coverage, adjusted IRS, t-SNE visualizations, and image-generation costs. These analyses provide useful information about the benefits and limitations of the method.

Critical weaknesses
  1. The core innovation and its independent contribution remain insufficiently validated. The authors claim that external knowledge helps identify plausible intra-class variations beyond the original few-shot samples. However, the current experiments mainly show improvements from the full KICA augmentation pipeline. They do not isolate the effects of external knowledge, prompt design, and editing strategy. Consequently, the experiments do not establish knowledge guidance as the primary driver of the reported gains.
  2. The reliability of generated images lacks direct validation. All generated images inherit the labels of their source images. The paper does not report a quantitative evaluation of class consistency. This omission raises concerns about potential label noise, particularly in fine-grained classification tasks.
  3. The experimental analysis does not fully support the conclusions. The paper does not clearly establish the contribution of individual attribute types in the multi-source knowledge bank. Some distribution metrics also fail to improve consistently. In addition, the main results lack information about statistical variation. These limitations weaken the interpretation of the proposed mechanism and the evidence supporting the conclusions.
Questions for authors

no

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

Related papers

Discussion (0)

Sign in to comment.