Path-based attribution methods like Integrated Gradients (IG) are widely used to interpret neural networks, yet imperceptible input perturbations can drastically alter their attribution maps while leaving model predictions unchanged. We introduce Entropy-Regularized Integrated Gradients (ERIG), a path-based attribution method that intrinsically attenuates gradient irregularities along the integration path using maximum-entropy softmax weighting with temperature parameter τ. We quantify the effect of the entropy-regularized weighting on attribution stability by using the Lipschitz properties of the model and its gradients. We decompose the attribution change under an input perturbation into a model-response term, whose bound is τ-independent, and a weight-variation term, whose bound decreases at rate O(1/(2τ)). Their ratio R(τ) characterizes the trade-off between gradient information and attribution stability. We further add a new dimension to attribution evaluation by adapting classical noise estimators into two metrics, LINV-C and LINV-P, that quantify attribution noise. Across eleven CNN and Transformer backbones on ImageNet and Oxford-IIIT Pet datasets, our experiments show that, relative to the next-best method, ERIG remains comparable on ranking-based fidelity metrics while improving localization by approximately 24% and attribution stability by approximately 47-fold on average, reaching up to four orders of magnitude improvement over IG, all at negligible additional computational cost.
Figures & tables
Figure 1: Entropy-Regularized Integrated Gradients (ERIG). An integration path Γ(t) connects the baseline xb to the input xa . At each path point x(k) , the indirect weighting scheme maps the raw gradient g=∇f to importance weights wi∗ through a softmax over the squared gradients g⊙g with temperature τ . Intuitively, the weights stay close to uniform unless the local sensitivity warrants concentration, and τ controls how readily they concentrate. Summing the weighted contributions wi∗⋅δf(k) over the path yields the final attribution ϕi .
Figure 2: Weight concentration along the integration path (100 ImageNet validation images, K=200 steps). Top, mean output fˉk and increment magnitude ∣δfk∣ ; bottom, mean Gini coefficient of the per-step weight (w(g)) maps. Only ERIG ( τ=3 ) suppresses noise over flat and saturated segments (NS) and enhances signal over the informative segment (SE).
Figure 3: MobileNetV4 Visualizations — More examples in A.3
Model
Method
ISC (↑)
AIC (↑)
DEL (↓)
ResNet101
BIG
0.3101
0.6348
0.4304
GIG
0.3812
0.5500
0.5765
IDGI
0.4923
0.6571
0.3211
IG
0.3951
0.6838
0.3884
ERIG (ours)
0.4893
0.6610
0.3270
SmoothGrad
0.1833
0.4011
0.5138
Table 1: ImageNet Fidelity: ERIG places first or second on every metric and backbone, with a +0.44% average lead.
Model
Method
IoU (↑)
PA (↑)
PF1 (↑)
ResNet101
BIG
0.0643
0.6386
0.1008
GIG
0.0450
0.6286
0.0718
IDGI
0.0212
0.6319
0.0354
IG
0.0782
0.6401
0.1200
ERIG (ours)
0.1617
0.6580
0.2419
SmoothGrad
0.1286
0.6498
0.1824
Table 2: ImageNet-S Localization: ERIG improves localization by ∼28% on average over the next-best method.
Model
Method
MS (↓)
LLC (↓)
RIS (↓)
ResNet101
BIG
0.1750
0.0078
12.498
GIG
0.3723
0.0166
88.373
IDGI
0.0100
0.0004
18.653
IG
0.6872
0.0307
30.295
ERIG (ours)
8.6e –05
3.8e –06
1.066
SmoothGrad
0.1391
0.0062
102.340
Table 3: ImageNet Stability: ERIG improves stability by ∼55× on average over the next-best method.
Model
Method
LINV-C (↑)
LINV-P (↑)
ResNet101
BIG
6.6490
6.6053
GIG
7.9654
7.9093
IDGI
9.4308
9.4017
IG
6.5869
6.4886
ERIG (ours)
16.8069
16.7781
SmoothGrad
10.1355
9.9179
Table 4: ImageNet Noise: ERIG attains the highest LINV on every backbone, 6.7 – 10.5 LINV above the next best.
Figure 4: MobileNetV4 Noise Histograms
Figure 5: Ablation of softmax temperature τ in ERIG across ResNet101, ViT-Base, and EVA02 models. All plots use a log-scale x-axis and a linear-scale y-axis, except stability, which also uses a log-scale y-axis.
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Additional visualizations of attribution maps for the MobileNetV4 architecture.
Figure 7: ISC distribution with inter-quartile range and full range across all architectures. Higher is better.
Figure 8: AIC distribution with inter-quartile range and full range across all architectures. Higher is better.
Figure 9: DEL distribution with inter-quartile range and full range across all architectures. Lower is better.
Figure 10: IoU distribution with inter-quartile range and full range across all architectures. Higher is better.
Figure 11: Pixel Accuracy distribution with inter-quartile range and full range across all architectures. Higher is better.
Figure 12: Pixel F1 distribution with inter-quartile range and full range across all architectures. Higher is better.
Figure 13: Max Sensitivity (MS) with inter-quartile range and full range across all architectures ( log10 scale). Lower is better.
Figure 14: Local Lipschitz Constant (LLC) with inter-quartile range and full range across all architectures ( log10 scale). Lower is better.
Figure 15: Relative Input Stability (RIS) with inter-quartile range and full range across all architectures ( log10 scale). Lower is better.
Figure 20
Figure 17: Mean total runtime as a function of the number of integration steps over 1000 examples. The left subplot shows IG, IDGI, and ERIG, which scale linearly and overlap considerably. The right subplot shows BIG and GIG on a separate scale due to their higher costs.
Table 22
Figure 18: Ablation of the softmax temperature τ in SmoothGrad+ERIG ensembles across MobileNetV4, ResNet101, and ViT-Base models. All plots use a log-scale x-axis and a linear-scale y-axis, except stability, which also uses a log-scale y-axis.
Table 24Table 25
Completeness
Correctness
Contrastivity
Model
Method
CSDC
PC
DC
Dist
Com.
SD
TS
ResNet-50
IG
0.920
0.940
0.880
0.810
0.862
0.590
0.980
IDGI
0.934
0.958
0.912
0.788
0.861
0.615
0.958
BIG
0.794
0.700
0.676
0.699
0.711
0.440
0.491
GIG
0.765
0.684
0.646
0.799
0.749
0.437
0.780
SmoothGrad
0.827
0.836
0.810
0.795
0.810
0.545
0.873
Appendix
Table 12: FunnyBirds Protocol: ERIG places first or second in all 21 cells, leading completeness by 3.4% and contrastivity by 2.1% .
Table 27Table 28
Noise
Complexity / Compressibility
Sparsity
Model
Method
LINV-C
LINV-P
H
Hhist
JPEG ratio
JPEG RMSE
GS
ResNet101
IG
6.587
6.489
0.9976
5.937
0.3957
3.345
0.6989
IDGI
9.431
9.402
0.8910
4.528
0.2790
2.673
0.8346
BIG
6.649
6.605
0.9978
5.674
0.3868
3.325
0.7304
GIG
7.965
7.909
0.9980
4.600
0.3272
2.998
0.8036
SmoothGrad
10.135
9.918
0.9968
6.675
0.4925
3.705
0.6000
Appendix
Table 21: LINV vs. Concentration: Compared measures disagree not only with LINV but also among themselves on the overall order.
Cells
Metric
H
Hhist
JPEG ratio
JPEG RMSE
GS
MS
LLC
RIS
All 36 cells
LINV-C
+0.66
+0.55
+0.57
+0.60
+0.36
+0.91
+0.92
+0.52
LINV-P
+0.65
+0.56
+0.60
+0.63
+0.38
+0.92
+0.93
+0.55
Without ERIG
LINV-C
+0.43
+0.28
+0.36
+0.38
+0.27
+0.84
+0.86
+0.17
LINV-P
+0.42
+0.31
+0.41
+0.42
+0.30
+0.87
+0.89
+0.23
Appendix
Table 22: LINV Correlations: LINV correlates at +0.91 to +0.93 with the stability metrics but at most +0.66 with any other compared measure. Even with ERIG excluded, the stability correlations stay at +0.86 to +0.89 while the rest collapse further.
Model
Method
AS (↓)
ALC (↓)
AIS (↓)
ResNet101
IG
2.0400
0.0811
117.301
IDGI
0.0334
0.0014
65.461
ERIG (ours)
1.3e–04
5.7e–06
1.033
EfficientNet
IG
1.2695
0.0331
65.008
IDGI
0.0254
7.4e–04
77.782
ERIG (ours)
9.0e–04
2.8e–05
14.239
Appendix
Table 23: ImageNet Adversarial Stability: ERIG is lowest on every backbone and metric, with a ∼90× improvement over the next-best method on average. Moreover, its adversarial AS and ALC stay below every baseline’s uniform MS and LLC respectively.
Figure 19: Schematic – Why the Signed Variant Cancels: ERIG’s per-step terms carry the sign of δf(k) , which is shared by every feature, so nothing cancels if f keeps increasing, and both on-object and background features accumulate alike. SERIG multiplies the weights by sign(gi) . This sign is usually stable on the object but alternating along the path for background, so only the background feature cancels toward zero in the aggregation (summation). Here, the hypothetical per-step weight magnitudes follow the movement of the model output along the path ( Fig. 2 ), small over the flat and saturated segments and large where the output changes rapidly.
We introduce path-sampled integrated gradients (PS-IG), a framework that generalizes feature attribution by computing the expected value over baselines sampled along the linear interpolation path. We prove that PS-IG is mathematically equivalent to path-weighted integrated gradients, provided the weighting function matches the cumulative distribution function of the sampling density. This equivalence allows the stochastic expectation to be evaluated via a deterministic Riemann sum, improving the error convergence rate from O(m−1/2) to O(m−1) for smooth models. Furthermore, we demonstrate analytically that PS-IG functions as a variance-reducing filter against gradient noise - strictly lowering attribution variance by a factor of 1/3 under uniform sampling - while preserving key axiomatic properties such as linearity and implementation invariance.
Firuz Kamalov, Fadi Thabtah, R. Sivaraj +1
Department of Computational Sciences, Canadian University Dubai, UAE · Abu Dhabi School of Management, Abu Dhabi, UAE · Department of Mathematics, Dr. B.R. Ambedkar National Institute of Technology, India
Feature attribution is central to diagnosing and trusting deep neural networks, and Integrated Gradients (IG) is widely used due to its axiomatic properties. However, IG can yield unreliable explanations when the integration path between a baseline and the input passes through regions with noisy gradients. While Guided Integrated Gradients reduces this sensitivity by adaptively updating low-gradient-magnitude features, input-space guidance still produces intermediate inputs that deviate from the data manifold. To address this limitation, we propose Manifold-Aligned Guided Integrated Gradients (MA-GIG), which constructs attribution paths in the latent space of a pre-trained variational autoencoder. By decoding intermediate latent states, MA-GIG biases the path toward the learned generative manifold and reduces exposure to implausible input-space regions. Through qualitative and quantitative evaluations, we demonstrate that MA-GIG produces faithful explanations by aggregating gradients on path features proximal to the input. Consequently, our method reduces off-manifold noise and outperforms prior path-based attribution methods across multiple datasets and classifiers. Our code is available at https://github.com/leekwoon/ma-gig/.
Soyeon Kim, Seongwoo Lim, Kyowoon Lee +1
Kim Jaechul Graduate School of AI, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea · INEEJI, Seongnam, Korea · KAIST InnoCORE LLM, KAIST, Daejeon, Korea
Integrated Gradients (IG) is a widely adopted feature attribution method that satisfies desirable axiomatic properties. However, the choice of integration path significantly affects the quality of attributions, and the standard straight-line path introduces all input features simultaneously, often accumulating noisy gradients along the way. To address this limitation, we propose Spectral Integrated Gradients, which constructs integration paths based on singular value decomposition (SVD) of the baseline-to-input difference. By progressively activating singular components from largest to smallest, SIG introduces global structure before fine-grained details, naturally following a coarse-to-fine progression. Through extensive evaluation across diverse image classification datasets, we demonstrate that SIG produces cleaner attribution maps with reduced noise and achieves improved quantitative performance compared to existing path-based attribution methods. Our code is available at https://github.com/leekwoon/sig/.
Soyeon Kim, Seongwoo Lim, Kyowoon Lee +1
Korea Advanced Institute of Science and Technology Seongnam, Gyeonggi-do, Republic of Korea · INEEJI Corp. Seongnam, Gyeonggi-do, Republic of Korea