Attribution Through the Decoder: Prediction-Space Explanations for Structured Outputs
Organizations: Duke University, Durham, NC
Abstract
Networks that predict angles, gaze directions, and rotations do not output these quantities directly: they emit raw coordinates that a fixed decoder, such as normalization, maps to the final prediction. Attribution methods are usually applied to the raw coordinates, so they can credit inputs for changes the decoder discards, such as rescaling a vector when only its direction matters. The resulting explanation can highlight inputs that have no effect on the prediction. We instead attribute through the decoder differential, which for normalization decoders reduces to a scaled tangent projection. We characterize invariance to output rescaling, show that every fixed smooth scalar target on a sphere has points of zero tangential sensitivity, and show that an endpoint-projected variant of Integrated Gradients can still attribute to scale-only inputs, whereas direct angular Integrated Gradients cannot. On synthetic tasks with analytically known relevance, ambient attribution assigns up to 72% of its mass to an irrelevant region. On MPIIGaze validation data, tangent Grad-CAM produces 44% greater prediction displacement under 10% pixel deletion. On PASCAL3D+ with a 6D rotation decoder, the full differential improves mean deletion displacement by 14.1 degrees over the ambient coordinate sum and by 13.8 degrees over a decoder-invariant scalar. Results for a two-polarity tangent LayerCAM variant are mixed: it exceeds ambient-sum LayerCAM on circular regression but no difference is detected on quaternion pose. These results separate two failure sources, radial contamination and scalarization loss. The approach uses existing attribution methods with decoder-aware targets, although some variants also require changes to attribution aggregation. An invariant attribution can still miss inputs that drive the prediction, so invariance and deletion performance should each be checked.
Figures & tables
| Method | via | ||
|---|---|---|---|
| Ambient-gradient-magnitude (Grad-CAM / HiResCAM / Saliency) | 0.428–0.520 | 0.281–0.508 | 0.747–0.757 |
| Tangent (all three) | 1.000 | 0.9996–0.9997 | 0.9998–0.9999 |
| Radial (all three) | 0.000 | 0.000–0.0003 | 0.0002–0.0005 |
| Circular / spherical LIME | 0.946 | 0.823 | 0.895 |
| Tangent-norm path integral | 1.000 | 0.9998 |
| Task / deletion | Ambient-sum | Decoder-invariant scalar | Full differential | vs. ambient / invariant scalar |
|---|---|---|---|---|
| Gaze / 10% | – | / – | ||
| Gaze / mean | – | / – | ||
| Quaternion pose / mean | – | / – | ||
| 6D pose / mean | / |
| Method | R18 | R50 | EffV2-S | ViT-B |
|---|---|---|---|---|
| Ambient-sum Grad-CAM | 2.10 | 0.94 | 0.65 | 1.10 |
| Tangent Grad-CAM | 1.58 | 1.32 | 1.94 | 0.73 |
| Ambient-sum LayerCAM | 5.06 | 4.45 | 3.53 | 0.75 |
| Two-polarity tangent LayerCAM | 5.24 | 4.53 | 4.33 | 1.44 |
| Scalar-angle LIME | 2.00 | 2.23 | 2.45 | 4.02 |
| Circular LIME | 2.32 | 2.36 | 2.68 | 4.78 |
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | 0.1 | 0.2 | 0.3 | 0.4 | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 | Mean |
|---|---|---|---|---|---|---|---|---|---|---|
| Tangent | 5.86 | 7.88 | 9.17 | 9.90 | 10.34 | 10.55 | 10.58 | 10.51 | 10.36 | 9.46 |
| Spherical occlusion | 7.86 | 8.88 | 9.42 | 9.68 | 9.79 | 9.80 | 9.69 | 9.61 | 9.65 | 9.37 |
| Spherical LIME | 4.60 | 6.29 | 7.53 | 8.43 | 8.98 | 9.43 | 9.70 | 9.94 | 10.07 | 8.33 |
| Spherical KernelSHAP | 4.38 | 5.79 | 6.92 | 7.88 | 8.77 | 9.36 | 9.77 | 10.00 | 9.97 | 8.09 |
| Ambient-sum | 4.21 | 5.62 | 6.53 | 7.10 | 7.68 | 8.41 | 9.30 | 10.05 | 10.39 | 7.70 |
| Endpoint-projected IG | 6.46 | 7.23 | 7.58 | 7.74 | 7.84 | 7.92 | 8.03 | 8.13 | 8.19 | 7.68 |
| Backbone | AIG (tan) | AIG (dir) | Std IG (best) | |
|---|---|---|---|---|
| ResNet18 | 6.166 | 6.157 | 6.164 | |
| ResNet50 | 7.078 | 7.065 | 7.064 | |
| EfficientNet | 6.333 | 6.360 | 6.344 | |
| ViT | 8.397 | 8.394 | 8.427 | |
| IG variant | Target / units | Mean abs err | Median ratio | |
| Direct | wrapped (rad) | 0.0186 | 1.008 |
| Method | Invariant | ResNet18 | ResNet50 | EfficientNetV2- | ViT- |
|---|---|---|---|---|---|
| Ambient-sum Grad-CAM | 2.10 | 0.94 | 0.65 | 1.10 | |
| Ambient-sum HiResCAM | 2.11 | 0.94 | 0.65 | 0.74 | |
| Tangent Grad-CAM | 1.58 (0.661) | 1.32 (0.493) | 1.94 (0.454) | 0.73 (0.519) | |
| Tangent HiResCAM (= tangent LayerCAM) | 1.58 (0.661) | 1.32 (0.493) | 1.94 (0.454) | 1.46 (0.535) | |
| Ambient-sum LayerCAM (rectified) | 5.06 | 4.45 | 3.53 | 0.75 | |
| Rectified tangent LayerCAM | 5.24 | 4.53 | 4.33 | 1.44 |
| Deletion (deg) | Seed-level | Bootstrap (per-image) | |||
|---|---|---|---|---|---|
| Fraction | Tangent | Ambient-sum | SD | [95% CI] | |
| 0.1 | 0.001 | ||||
| 0.2 | |||||
| 0.3 | 0.001 | ||||
| 0.4 | 0.003 | ||||
| 0.5 | 0.013 |
| Backbone | Test mean | Test med. | |||
|---|---|---|---|---|---|
| ResNet18 | 1.48 | 0.81 | 34.5% | 60.7% | 82.8% |
| ResNet50 | 1.33 | 0.69 | 39.5% | 62.9% | 86.5% |
| EfficientNetV2- | 1.30 | 0.64 | 41.1% | 65.5% | 83.8% |
| ViT- | 1.42 | 0.73 | 39.0% | 60.7% | 84.4% |
| Random guess | 6.00 | ||||
| Constant (mean time) | 2.32 |
| Method | ||
|---|---|---|
| Model angular error | ||
| Ambient-gradient-magnitude Grad-CAM | 0.804 (0.081) | 0.988 (0.006) |
| Tangent Grad-CAM | 0.802 (0.079) | 0.988 (0.007) |
| Radial Grad-CAM | 0.807 (0.083) | 0.985 (0.014) |
| Ambient-gradient-magnitude HiResCAM | 0.663 (0.059) | 0.721 (0.020) |
| Tangent HiResCAM | 0.667 (0.060) | 0.722 (0.022) |
| Kernel | Surrogate | Deletion AUC (h) | std |
|---|---|---|---|
| Euclidean | scalar | 2.501 | 1.978 |
| von Mises | scalar | 2.362 | 1.952 |
| Combined | scalar | 2.387 | 1.892 |
| Euclidean | circular | 2.573 | 1.764 |
| von Mises | circular | 2.575 | 1.847 |
| Combined | circular | 2.590 | 1.854 |
| Backbone | Amb. | Tangent |
|---|---|---|
| ResNet18 | 0.002 | 0.000 |
| ResNet50 | 0.000 | 0.000 |
| EfficientNetV2- | 0.000 | 0.000 |
| ViT- | 0.359 | 0.725 |
| Backbone | Baseline (A) | Radial penalty (C) | Tangential penalty (F) |
|---|---|---|---|
| ResNet18 | .058 | .107 | .052 |
| ResNet50 | .060 | .027 | .022 |
| ViT- | .066 | .023 | .030 |
| Predictor | Pooled | Condition-mean | Seed , | Permutation |
|---|---|---|---|---|
| gap | , 0.0017 | ( ) | ||
| energy gap | , 0.0075 | ( ) | ||
| map overlap gap | , 0.0062 | ( ) | ||
| probe gap | , 0.0049 | ( ) |
| Attribution target | Decoder-invariant? | Collapse-free? | Mean AUC (deg) |
|---|---|---|---|
| Ambient sum over raw | no | no | 48.9 |
| Sum over decoded entries | yes | no | 49.2 |
| Tangent differential through | yes | yes | 63.0 |
| Target | Deletion AUC (h) | std |
|---|---|---|
| cos only | 2.946 | 1.514 |
| Output norm | 2.654 | 2.195 |
| 2.571 | 1.663 | |
| sin only | 1.981 | 1.618 |
| Decoded angle | 1.655 | 1.205 |
| Tangent projection | 1.643 | 1.200 |
| Arm | Description | Val err (h) | 10-seed | |
|---|---|---|---|---|
| A | Baseline (tanh head, no penalty) | 1.076 | 1.021 | 0.200 |
| -normalized head | 1.069 | |||
| Radial penalty, | 1.087 | 1.019 | 0.167 | |
| Radial penalty, | 1.780 | 0.143 | ||
| Generic grad-norm penalty | 1.083 | 0.194 | ||
| Tangential penalty, | 1.131 | 1.025 | 0.261 |
| Deletion (deg) | Seed-level | Bootstrap | |||
|---|---|---|---|---|---|
| Fraction | Tangent | Ambient-sum | SD | [95% CI] | |
| 0.1 | 0.001 | ||||
| 0.4 | 0.003 | ||||
| 0.7 | 0.870 | ||||
| 0.9 | 0.416 | ||||
| Mean | 0.017 |
| Fraction | Tangent | Ambient-sum | SD | [95% CI] | |
|---|---|---|---|---|---|
| 0.1 | 0.015 | ||||
| 0.4 | 0.024 | ||||
| 0.7 | 0.029 | ||||
| Mean | 0.013 |
Feature Attribution
Feature attribution estimates which input features or internal components contribute to a model’s predictions.
See all 253 papers