PaperlayerView exact revision on arXiv

Source-linked AI summary

Understanding the Impact of Model Pruning on Long-Tail Forgetting and Explanation Reliability in Medical Imaging

Nazish Khalid, Tausifa Jan Saleem, Amal Saqib, Donald C. Wunsch, Mohammad Yaqub

arXiv:2609.07803v1cs.AI

TL;DR

The paper examines whether pruning causes long-tail forgetting and undermines explanation reliability in medical imaging, where rare conditions are underrepresented. It systematically evaluates multiple datasets, architectures, pruning methods, sparsity levels, and explanation measures, finding that class frequency primarily shapes predictive degradation while pruning strategy primarily shapes explanation reliability. The results support evaluating compressed medical models with both class-aware and explanation-aware measures.

  • Problem

    It remains unclear how pruning affects rare classes in long-tailed medical datasets and whether pruned models preserve reliable explanations.

  • Method

    The study evaluates four pruning strategies across two long-tailed medical imaging datasets, two CNN architectures, sparsity up to 95%, and predictive, stability, and faithfulness measures.

  • Results

    Lower-frequency classes generally degrade earlier and more severely, while explanation stability and faithfulness depend mainly on pruning strategy, with gradient-informed methods performing better under aggressive compression.

  • Takeaways & Limitations

    Compressed medical AI should be evaluated beyond aggregate performance using class-aware and explanation-aware measures, with moderate sparsity offering a practical balance among competing objectives.

Abstract

from arXiv · show

Model pruning is widely used to compress deep neural networks, reducing memory and computational requirements with minimal impact on aggregate performance. However, its effect on model behavior remains poorly understood, particularly for long-tailed medical datasets where rare but clinically important conditions are underrepresented. Furthermore, it remains unclear whether pruned models preserve reliable explanations of their predictions. To address this gap, we present a systematic study of long-tail forgetting and explanation reliability under model pruning. Across two long-tailed medical imaging datasets, two CNN architectures, four pruning methods, and sparsity levels up to 95\%, we evaluate predictive performance, explanation stability, and explanation faithfulness. Our results show that predictive performance exhibits a strong frequency-dependent trend, with lower-frequency classes generally experiencing earlier and larger degradation than higher-frequency classes. In contrast, explanation stability and faithfulness are influenced primarily by the pruning strategy, with gradient-informed methods preserving explanation reliability more effectively under aggressive compression. Qualitative and mechanistic analyses further indicate that explanation degradation is primarily associated with the collapse of class-discriminative gradients rather than the disappearance of feature activations. These findings suggest that model compression should be evaluated beyond aggregate performance. Incorporating class-aware and explanation-aware evaluation reveals failure modes that would otherwise remain hidden, while moderate sparsity levels provide a practical balance between compression, predictive performance, and explanation reliability.

I. INTRODUCTION

Model pruning reduces computational demands but can alter class-specific behavior and explanation reliability, especially in long-tailed medical imaging. This study addresses whether these effects depend on pruning strategy, class frequency, and explanation dimension.

  • Pruning reduces memory and computational costs by removing redundant weights, often with minimal aggregate-performance impact.
  • Aggregate metrics can mask clinically important behavioral changes, including failures on medically significant cases.
  • Rare medical conditions are more vulnerable to pruning than common conditions because long-tailed datasets contain far fewer examples for those classes.
  • Unreliable explanations could make compressed models highlight different image regions from dense models despite acceptable aggregate accuracy.
  • The study asks whether long-tail forgetting generalizes across pruning methods, how strategies affect frequency groups, and how pruning changes explanation stability, faithfulness, and predictive performance.
  • Experiments compare four pruning strategies across two long-tailed medical datasets, two CNN architectures, and sparsity levels from 0% to 95%.
  • The paper jointly evaluates attribution stability and faithfulness alongside predictive performance to study explanation reliability under compression.

C. Impact of Pruning

The paper combines prior evidence on pruning-induced forgetting with a cross-domain evaluation of datasets, architectures, and training formulations. Its design targets whether pruning behavior generalizes across medical imaging settings and model types.

  • Prior work and study gap: Prior work found that pruning can disproportionately affect some classes, with rare diseases more vulnerable than common diseases.
  • Prior work and study gap: Existing medical studies mainly examine magnitude pruning or evaluate predictive performance and explainability separately.
  • Study design: The study compares L1, SNIP, GraSP, and Random pruning across two long-tailed datasets and two CNN architectures.
  • Study design: The datasets span thoracic radiology and dermatology, providing distinct imaging modalities, task formulations, and class-imbalance conditions.
  • Datasets: NIH-CXR-LT contains 20 thoracic-disease and No Finding classes grouped into head, mid, and tail categories by training frequency.
  • Datasets: ISIC-2019 contains eight skin-lesion categories, excludes the Unknown test class, and uses an 85%/15% stratified training-validation split.
  • Task formulation: Both datasets use one-vs-rest classification with sigmoid outputs and binary cross-entropy, while the study preserves natural class imbalance without reweighting or resampling.
  • Architectures: ResNet-50 and DenseNet-121 provide complementary residual and dense-connectivity settings with approximately 25.6M and 8.0M parameters, respectively.

C. Pruning Strategies

The study compares four pruning strategies across sparsity levels, finding distinct compression behaviors and frequency-dependent effects on class-wise performance.

  • Pruning methods: SNIP estimates weight importance from first-order sensitivity to training loss at initialization and removes the least salient connections.Its design seeks to preserve loss-sensitive connections under high sparsity.
  • Pruning methods: GraSP estimates importance using second-order information to preserve gradient flow during training.The method accounts for pruning effects on optimization dynamics.
  • Predictive performance: SNIP shows the most graceful mAP degradation, while L1 pruning collapses at high sparsity and Random pruning degrades rapidly even at low sparsity.The comparison spans NIH-CXR-LT and ISIC-2019 with ResNet-50 and DenseNet-121 backbones.
  • Long-tail effects: Lower-frequency classes generally experience earlier and larger AP degradation than higher-frequency classes across pruning strategies.Positive correlations between class frequency and pruning robustness quantify this frequency-dependent pattern.

D. Training Protocol

The training protocol uses standardized optimization, preprocessing, augmentation, initialization, and repeated evaluation across the pruning experiments.

  • Optimization: Models are optimized with Adam at a learning rate of 1 × 10^-4 for up to 60 epochs, using early stopping with patience 15.Early stopping is based on validation performance.
  • Data processing: Input images are resized to 224×224, normalized with ImageNet statistics, and augmented using horizontal flips and rotations of ±10°.These preprocessing choices are applied within the shared experimental protocol.
  • Implementation: All models use ImageNet-pretrained weights and are implemented in PyTorch, with GradCAM explanations generated using Captum.The protocol therefore evaluates both model performance and explanation outputs within a common software setup.
  • Evaluation: Each experiment is repeated over three random seeds, producing 960 evaluated experimental configurations.The repeated-seed design supports reporting median or aggregated behavior across runs.
  • Evaluation: At 95% sparsity, relative AP change is evaluated against the logarithm of class frequency to characterize long-tail degradation.The analysis examines whether lower-frequency classes undergo more severe performance loss.

IV. EVALUATION FRAMEWORK

The evaluation combines predictive and explanation-reliability measures to assess pruning under long-tailed medical imaging conditions. It uses mAP and per-class AP alongside Grad-CAM stability and faithfulness measures.

  • Explanation reliability: Explanation reliability is assessed with Grad-CAM using stability relative to the dense model and faithfulness of highlighted regions to predictions.These complementary dimensions evaluate both explanation consistency and whether salient regions contribute to the model’s output.
  • Predictive performance: Predictive performance is measured with mean Average Precision and per-class Average Precision, with per-class AP stratified by frequency to quantify long-tail forgetting.AP is preferred over AUROC because it is more sensitive to severe class imbalance and better reflects rare-class performance.
  • Explanation stability: Stability IoU measures spatial agreement between the top 20% most salient pixels in dense and pruned Grad-CAM maps.Higher Stability IoU indicates greater agreement between the two explanation maps.

B. Explanation Reliability

Explanation reliability is evaluated by comparing pruned explanations with dense-model explanations and by testing whether salient regions affect predictions. Stability IoU captures spatial consistency, while AOPC captures explanatory contribution.

  • Grad-CAM: Grad-CAM generates attribution maps for an input image and target class using class-specific importance weights and the class score.The supplied passages identify the input image, target class, importance weight, and class score as components of the attribution process.
  • Explanation stability: Stability IoU compares the top 20% salient pixels in dense and pruned Grad-CAM maps using their intersection-over-union.The metric is averaged over test images and positive class labels.
  • Explanation faithfulness: AOPC measures explanation faithfulness by tracking prediction-confidence reduction as increasingly salient image regions are masked.Higher AOPC indicates that highlighted regions contribute more strongly to the model’s prediction.

3) Explanation Faithfulness:

Explanation faithfulness remains comparatively robust for gradient-informed pruning but varies substantially by pruning strategy. Predictive performance and explanation stability show stronger class-frequency and high-sparsity effects than faithfulness does.

  • Explanation faithfulness: AOPC measures faithfulness through prediction-confidence reduction as salient regions are progressively masked.Higher AOPC indicates that the highlighted regions contribute more strongly to the prediction; samples with p(0)_c < 0.01 are excluded and T = 20.
  • Evaluation procedure: Explainability metrics use stratified subsets of 500 test images per dataset while preserving the original class-frequency distribution.Grad-CAM maps are normalized to [0, 1], computed for positive labels, and averaged over three random seeds.
  • Predictive performance: SNIP preserves class-wise predictive performance most effectively, while Random pruning degrades rapidly across all frequency groups even at low sparsity.L1 shows a pronounced high-sparsity collapse, and GraSP incurs an early penalty followed by comparatively stable degradation.
  • Explanation stability: Explanation stability is primarily method-dependent, with no consistent head–mid–tail ordering across class frequencies.SNIP maintains relatively stable correspondence across frequency groups, while L1 shows high-sparsity collapse and GraSP shifts early before stabilizing.
  • Explanation faithfulness: SNIP and GraSP preserve explanation faithfulness across a wide sparsity range, whereas L1 collapses at high sparsity and Random pruning degrades rapidly at low sparsity.The results indicate stronger faithfulness retention for gradient-informed methods under aggressive compression.

B. Explanation Stability Under Pruning

Explanation reliability under pruning depends more on pruning strategy than class frequency. Gradient-informed methods, especially SNIP, preserve stability and faithfulness better under aggressive compression, while degradation is linked to collapsed class-discriminative gradients.

  • Stability: SNIP maintains the highest spatial agreement with the dense model at extreme sparsity, whereas Random pruning rapidly loses explanation consistency.GraSP shifts early but then remains relatively stable; L1 collapses sharply at high sparsity.
  • Stability: Explanation stability shows no consistent head–mid–tail ordering and is primarily determined by the pruning strategy.SNIP remains relatively stable across representative frequency groups, while Random degrades across them.
  • Faithfulness: SNIP and GraSP maintain relatively stable AOPC values across sparsity, indicating that highlighted regions continue contributing substantially to predictions.Faithfulness is more robust to compression than predictive performance for these gradient-informed methods.
  • Faithfulness: Random pruning performs worst, with AOPC approaching zero shortly after pruning begins, while L1 collapses at high sparsity.Per-class faithfulness differences are primarily method-dependent rather than consistently associated with class frequency.
  • Qualitative analysis: At 95% sparsity, SNIP most consistently preserves lesion-centered localization, whereas Random often produces diffuse, edge-dominated, or nearly blank maps.These qualitative trends hold across both datasets and architectures.
  • Qualitative analysis: SNIP explanations change gradually with increasing sparsity, retaining diagnostically relevant regions even at high sparsity.This contrasts with the abrupt degradation observed under L1 pruning.
  • Mechanism: L1 can retain substantial non-zero activations despite severely degraded explanations, while L1 and Random shift class-discriminative gradients toward near-zero values.Because Grad-CAM weights feature maps using these gradients, gradient collapse provides a plausible mechanism for explanation degradation.

VI. DISCUSSION

Pruning produces systematic long-tail forgetting across methods, with lower-frequency classes generally degrading earlier and more severely. Gradient-informed pruning, particularly SNIP, preserves class-wise performance better under aggressive compression.

  • Long-tail forgetting: Long-tail forgetting occurs across datasets, architectures, and pruning methods, with lower-frequency classes generally experiencing earlier and larger performance degradation.The pattern is systematic despite deviations by individual classes.
  • Long-tail forgetting: Frequent diseases may develop richer and more redundant feature representations, whereas rare diseases may rely on fewer discriminative pathways.The paper presents this as a hypothesis explaining the frequency-dependent pruning sensitivity.
  • Pruning strategy: Gradient-informed approaches, particularly SNIP, preserve class-wise performance substantially better than magnitude-based and random pruning.Retaining optimization-sensitive connections may help maintain minority-class representations under aggressive compression.
  • Mechanistic interpretation: At 95% sparsity, Grad-CAM gradient and activation fractions quantify the share of target-layer signals falling below near-zero thresholds.Grad % uses 10^-4 for gradients and Act % uses 0.01 for activations; higher fractions indicate signal collapse.

B. Explanation Stability and Faithfulness Provide Complementary Insights

Explanation stability and faithfulness capture complementary aspects of pruned-model behavior and respond differently from predictive performance. These findings support class-aware and explanation-aware evaluation, with moderate sparsity offering a practical operating regime, while generalization beyond CNNs remains open.

  • Complementary explanation metrics: Explanation stability and faithfulness are primarily method-dependent rather than consistently ordered by class frequency, unlike predictive performance.The two metrics therefore capture complementary aspects of explanation reliability.
  • Mechanism: L1 can preserve activation magnitudes while producing severely degraded attribution maps, and both L1 and Random exhibit collapsed class-discriminative gradients.This supports gradients, rather than feature activation presence alone, as a key determinant of Grad-CAM reliability.
  • Implications: Preserving predictive performance alone is insufficient to ensure reliable explanations after compression.Explanation quality depends critically on preserving class-discriminative gradients.
  • Implications: Compressed medical AI should be evaluated with class-aware and explanation-aware protocols rather than aggregate performance alone.Such evaluation better reflects robustness in safety-critical clinical applications.
  • Overall conclusion: Across four pruning strategies, two architectures, and two long-tailed datasets, gradient-informed pruning, particularly SNIP, provides the most favorable balance between predictive performance and explanation reliability.The study links aggressive-pruning explanation degradation primarily to collapsed class-discriminative gradients rather than disappearing high-level activations.
  • Deployment: Moderate sparsity consistently provides a practical operating regime that preserves both predictive performance and explanation reliability.
  • Scope boundary: The study is limited to CNN architectures, so whether the observed pruning behaviors generalize to transformer-based models remains unresolved.The authors also identify structured pruning, quantization, knowledge distillation, and broader imaging tasks as future directions.
Loading 2609.07803v1…