Attribution methods improved for neural networks predicting angles and rotations

Attribution Through the Decoder: Prediction-Space Explanations for Structured Outputs

Computer Vision and Pattern RecognitionArtificial Intelligence

Summary

Some neural networks predict directions and rotations, but they do this indirectly by giving raw coordinates, which are then converted by a fixed rule to the final answer. Usually, explanations of these predictions look only at the raw coordinates, which can wrongly highlight inputs that don’t actually affect the final prediction. The authors show a better way to explain predictions by including how the fixed rule (decoder) changes the output, making explanations more accurate. They test this on tasks like gaze direction and 3D rotations and show it improves how well the explanations match what really influences the prediction.

What this means in practice

  • For computer vision engineers: Improve explanations for models predicting gaze, rotation, or angles by attributing through decoders that map raw outputs to final predictions.
  • For robotics developers: Enhance interpretation of neural networks that estimate 3D object orientation by using decoder-aware attribution for more accurate input relevance assessment.

Authors

Zaer Bin Zaqy, Brinnae Bent

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.