Source-linked AI summary
Towards Better Explanations of Class Activation Mapping
Hyungsik Jung, Youngrock Oh
TL;DR
Most CAM methods lack a clear theoretical basis for choosing activation-map coefficients. This paper formulates coefficient selection as additive feature attribution, using SHAP values as a unified solution and introducing LIFT-CAM to approximate them. LIFT-CAM provides qualitatively enhanced explanations and state-of-the-art results on various quantitative evaluation metrics.
Problem
Most CAM methods assign activation-map coefficients heuristically without a clear theoretical basis or rigorous standards for a good explanation model.
Source: §1 · p. 1Method
The paper models activation-map presence with binary variables, uses SHAP values as coefficients, and introduces DeepLIFT-based LIFT-CAM to approximate them.
Source: §1 · p. 1, §3.3 · p. 3Results
LIFT-CAM provides qualitatively enhanced visual explanations and achieves state-of-the-art results on various quantitative evaluation metrics.
Source: §5 · p. 8Takeaways & Limitations
The framework treats SHAP values as a unified solution for the explanation model and also unifies Ablation-CAM within the framework.
Source: §5 · p. 8LIFT-CAM sets reference values for absent activation neurons to zero, aligning them with SHAP.
Source: §3.4 · p. 4
Abstract
from arXiv · showhide
Increasing demands for understanding the internal behavior of convolutional neural networks (CNNs) have led to remarkable improvements in explanation methods. Particularly, several class activation mapping (CAM) based methods, which generate visual explanation maps by a linear combination of activation maps from CNNs, have been proposed. However, the majority of the methods lack a clear theoretical basis on how they assign the coefficients of the linear combination. In this paper, we revisit the intrinsic linearity of CAM with respect to the activation maps; we construct an explanation model of CNN as a linear function of binary variables that denote the existence of the corresponding activation maps. With this approach, the explanation model can be determined by additive feature attribution methods in an analytic manner. We then demonstrate the adequacy of SHAP values, which is a unique solution for the explanation model with a set of desirable properties, as the coefficients of CAM. Since the exact SHAP values are unattainable, we introduce an efficient approximation method, LIFT-CAM, based on DeepLIFT. Our proposed LIFT-CAM can estimate the SHAP values of the activation maps with high speed and accuracy. Furthermore, it greatly outperforms other previous CAM-based methods in both qualitative and quantitative aspects.
1. Introduction
CAM methods combine activation maps, but most assign their coefficients heuristically without a clear theoretical basis. The paper reframes coefficient selection through additive feature attribution, identifies SHAP values as a solution, and introduces LIFT-CAM, which outperforms previous CAMs qualitatively and quantitatively.
1. Introduction
- The paper proposes determining plausible CAM explanations by finding a reliable solution for the explanation model with additive feature attribution methods.The framework can also reinterpret Ablation-CAM.Source: §1 · p. 1
- The paper presents SHAP values of activation maps as a unified solution for its explanation framework.Their benefits for generating faithful visual explanations are verified.Source: §1 · pp. 1–2
- LIFT-CAM estimates activation-map SHAP values using DeepLIFT and outperforms previous CAMs qualitatively and quantitatively.The method uses a single backward propagation.Source: §1 · p. 2
2. Related Work
The related work reviews CAM coefficient strategies, the SHAP framework for additive feature attribution, and DeepLIFT's contribution-based approach. It contrasts gradient-based and gradient-free CAMs and describes how DeepLIFT relates to SHAP values.
2.1. Class Activation Mapping
- CAMs linearly combine activation maps; Grad-CAM averages gradients, while Grad-CAM++ emphasizes positive influences using higher-order derivatives.Gradient saturation can make raw gradients fail to localize relevant regions; gradient-free Score-CAM and Ablation-CAM instead use predictions, but require N_l forward propagations.Source: §2.1 · p. 2
2.2. SHapley Additive exPlanations
- SHAP expresses an explanation as a baseline plus feature attributions for binary feature-presence variables, and uniquely satisfies local accuracy, missingness, and consistency.Its values weight output changes when a feature is removed across feature subsets.Source: §2.2 · p. 2
2.3. Deep Learning Important FeaTures
- DeepLIFT propagates differences from reference activations to assign input contributions, alleviating gradient saturation.Its summation-to-delta property makes it an additive feature attribution method, and it efficiently approximates SHAP values while satisfying local accuracy and missingness.Source: §2.3 · pp. 2–3
3. Methodology
The methodology reframes CAM coefficient selection as an additive feature-attribution problem over binary variables representing activation maps. It adopts SHAP values as a unified solution and introduces LIFT-CAM as an efficient approximation, while interpreting Ablation-CAM within the same framework.
3. Methodology
- The methodology’s central approach is to determine CAM coefficients analytically through additive feature attribution applied to a linear explanation model.The proposed framework reinterprets coefficient selection as finding a reliable solution for that model.Source: §3 · p. 3, §3.2 · p. 3
3.1. Problem formulation of CAM
- CAM’s explanation map is linear in fixed activation maps, so its explanatory quality depends on choosing coefficients that reliably explain the target output.The coefficients represent the activation maps’ importance scores.Source: §3.1 · p. 3
3.2. Proposed framework
- The framework represents activation-map presence with binary variables and models the explanation as a linear function of those variables.Each activation map is treated as an individual feature; additive feature-attribution methods determine the model’s coefficients.Source: §3.2 · p. 3
3.3. SHAP values of activation maps
- SHAP values provide a unified coefficient solution by averaging each activation map’s marginal prediction differences across all feature orderings.A subset of orderings can be used to estimate the values and reduce computational burden.Source: §3.3 · p. 3
3.4. Efficient approximation: LIFT-CAM
- LIFT-CAM approximates SHAP values by propagating DeepLIFT contributions from the target output to activation maps and summing contributions within each map.It uses a single backward pass; for linear F, it can obtain exact SHAP values, while few nonlinearities in F support high-precision approximation.Source: §3.4 · pp. 4–5
3.5. Rethinking Ablation-CAM
- Ablation-CAM’s marginal-difference coefficients can be viewed as another approximation to SHAP values, but require N_l forward simulations.It does not satisfy SHAP local accuracy, a mismatch associated with less precise approximations and less reliable explanations than LIFT-CAM.Source: §3.5 · p. 5
4. Experiments and Results
Experiments validate SHAP-based coefficients, assess LIFT-CAM’s approximation and compare its explanations across object recognition and VQA. LIFT-CAM performs strongly across faithfulness, localization, and visual comparisons, with one stated exception on COCO.
4.1. Validation of SHAP values
- SHAP-CAM100 performs best across the reported faithfulness cases, while increasing the number of orderings improves IC and ADD and lowers AD.The table evaluates SHAP-CAM1, SHAP-CAM10, and SHAP-CAM100, averaging each case over 10 simulations.Source: §4.1 · p. 6, Table 1 · p. 6
4.2. Approximation performance of LIFT-CAM
- LIFT-CAM has the highest coefficient cosine similarities to SHAP-CAM10k across datasets, exceeding 0.9.Ablation-CAM also has high similarities but falls behind LIFT-CAM due to dissatisfaction of local accuracy.Source: §4.2 · p. 6, Table 2 · p. 6
4.3. Performance evaluation of LIFT-CAM
- The evaluation compares LIFT-CAM with previous state-of-the-art saliency methods on visualization, faithfulness, and localization.The experiments also apply LIFT-CAM to VQA.Source: §4 · p. 5, §4.3 · p. 6
4.3.1 Qualitative evaluation
- LIFT-CAM produces visually interpretable maps across the examples, reducing pixel noise and locating object regions, including multiple apples.The examples include a notebook, a sheep, and apples; Figure 3 compares saliency-method maps.Source: §4.3.1 · p. 6, Figure 3 · p. 7
4.3.2 Faithfulness evaluation
- LIFT-CAM leads most object-recognition IC, AD, and ADD cases and has the highest insertion and lowest deletion AUC; Ablation-CAM narrowly leads COCO ADD.The COCO ADD difference is described as negligible; LIFT-CAM requires only a single backward pass to calculate coefficients.Source: §4.3.2 · pp. 6–7, Table 3 · p. 7, Table 4 · p. 7
4.3.3 Localization evaluation
- LIFT-CAM has the highest proportion of explanation-map energy inside bounding boxes among the compared methods.The paper interprets this as a compact map focused on essential image regions without trivial noise.Source: §4.3.3 · p. 8, Table 5 · p. 8
4.4. Application to VQA
- On VQA, LIFT-CAM outperforms Grad-CAM on IC, AD, and ADD and highlights image regions more relevant to the question-answer pairs.The comparison evaluates the metrics with fixed question embeddings.Source: §4.4 · p. 8
5. Conclusion
The conclusion presents an analytic framework for determining CAM coefficients and reports that LIFT-CAM approximates SHAP values with one backward pass, improving visual explanations and achieving state-of-the-art quantitative results.
5. Conclusion
- The framework determines CAM coefficients by optimizing a linear explanation model with additive feature attribution methods.It also unifies Ablation-CAM within the framework.Source: §5 · p. 8
- Figure 4 presents Grad-CAM and LIFT-CAM visual explanation maps for VQA.Source: Figure 4 · p. 8
- LIFT-CAM approximates SHAP values for activation maps with one backward pass, producing qualitatively enhanced explanations and state-of-the-art results on quantitative metrics.Table 6 covers faithfulness evaluation for VQA; its caption specifies that higher is better for IC and ADD, and lower for AD.Source: §5 · p. 8, Table 6 · p. 8