acceptodds
Under review as a conference paper at ICLR 2027

Explanation-Guided Active Learning by Exploring Diverse Decision Rationales

Abstract

Skeleton-based action recognition relies on large-scale annotations, motivating active learning to identify informative samples under limited labeling budgets. Existing acquisition strategies typically quantify informativeness using predictive uncertainty, feature-space coverage, or gradient-based sensitivity, without considering the evidence underlying model decisions. Samples supported by similar spatio-temporal evidence may provide redundant supervision even when they are distant in prediction, feature, or gradient space. Conversely, samples with similar representations may exhibit distinct decision patterns and thus provide complementary supervision. We propose an explanation-guided active learning framework that selects samples based on their decision evidence. Using Layer-wise Relevance Propagation, we identify prediction-preserving spatio-temporal evidence and encode it into fixed-dimensional explanation representations. Batch acquisition then improves coverage in this explanation space, promoting diversity in decision evidence beyond predictions, features, or gradients. We further bound the generalization risk by the explanation-space covering radius, optimization error, and explanation reconstruction error, providing theoretical support for our criterion. Experiments on two skeleton-based action recognition datasets consistently outperform existing active learning baselines.

Primary area:
interpretability and explainable AI
Keywords:
Active Learning, Explanations, Layer-wise Relevance Propagation

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 2 reviews, above 9% of the others.
Reviewer_ju8jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an explanation-guided active learning method for skeleton-based action recognition tasks. The method extracts skeleton spatiotemporal features based on CTR-GCN and uses LRP to obtain joint explanation representations; it optimizes the KL divergence objective function to select key joints and key frames, so that the retained feature information can reconstruct the original prediction distribution. The method achieves good results on two skeleton-based action recognition datasets.

Critical strengths

This paper introduces the spatiotemporal explanatory evidence underlying model predictions into active learning. It employs Layer-wise Relevance Propagation (LRP) and KL divergence to construct explanation vectors. Compared with conventional methods, the proposed framework offers a new perspective on sample selection and has potential research value.

Critical weaknesses
  1. Lack of computational cost analysis and ablation studies. The proposed method requires LRP computation and KL divergence optimization for sample selection, which may introduce substantial computational overhead. However, the authors do not provide a corresponding analysis of computational complexity or runtime costs. Additionally, ablation studies are missing to evaluate the individual contributions of the key components.

  2. Insufficient comparison with recent state-of-the-art methods. Most of the baseline methods considered in this paper are relatively outdated, and systematic comparisons with recent active learning methods proposed in the past two years are lacking.

Questions for authors
  1. Could the authors provide a comparison of the proposed method with the baselines in terms of runtime and computational cost? Additionally, could the authors conduct ablation studies to evaluate the individual contributions of LRP, KL divergence optimization, and explanation-space-based sample selection to the overall performance improvements?

  2. Could the authors explain why active learning methods proposed in the past two years were not included as baselines? Are there any experimental results comparing the proposed method with these recent approaches to further validate its competitiveness?

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

The paper tackles an interesting problem of active learning for human action recognition using skeleton videos. The paper addresses an interesting problem and presents a theoretically motivated approach to incorporating explanation diversity into active learning for skeleton-based action recognition. The technical formulation, including the extension of LRP to CTR-GCN, is a positive aspect of the work.

Critical strengths
  1. The paper proposes an interesting combination of Layer-wise Relevance Propagation (LRP), CoreSet-style coverage for active learning, and skeleton-based action selection. The overall motivation is clearly presented. In fact, reading the introduction of the paper is very enjoyable.
  2. The extension of LRP to the CTR-GCN architecture appears technically sound and is appropriately derived. This provides a reasonable foundation for incorporating explanation-based information into the active learning framework.
Critical weaknesses

At the high level, major concerns include using KL for measuring model prediction- Other acquisition strategies such as EPIG may be used. The architecture seems to unnecessarily complicated due to the many blocks being used. Further, does explanation-space diversity imply better learning? Many of these questions are not addressed.

  1. The proposed approach assumes that diversity in the spatio-temporal locations of prediction-preserving evidence corresponds to diversity in useful supervision. However, the final explanation representation retains only the normalized selection frequencies of joints across temporal stages, discarding class-specific relevance. What theoretical or empirical evidence establishes that distances between these representations reflect differences in label-relevant decision mechanisms, rather than merely differences in attribution patterns?
  2. Insufficient empirical validation: Although the proposed method is theoretically motivated, the experimental evaluation is not sufficient to substantiate the paper's main claims. In particular, the experiments do not convincingly establish the effectiveness of explanation-guided sample selection over existing approaches.
  3. Lack of comparison with relevant domain-specific baselines: The experimental evaluation primarily considers standard active learning baselines. However, the paper does not compare against existing active learning approaches specifically designed for skeleton-based action recognition, such as AL-SAR.
  4. Missing computational efficiency and statistical reliability analysis: The paper does not report runtime comparisons with existing baselines. Since the proposed framework introduces additional explanation computation and selection mechanisms, evaluating the associated computational overhead is important. Furthermore, the absence of results across multiple independent runs, including standard deviations or confidence intervals, makes it difficult to assess the robustness, consistency, and statistical significance of the reported improvements.
  5. Lack of ablation studies: The paper does not provide a dedicated ablation study to isolate the contributions of its key components, namely LRP-based explanations, KL-based selection, and explanation-space coverage. Without these experiments, it is unclear which components are responsible for the observed improvements and whether their combination provides meaningful benefits over simpler alternatives.
Questions for authors

Clarify the points in wekanesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.