← News

ResearchSources: arXiv · 18 Nov 2026

Researchers propose attributing model predictions through the decoder, not raw coordinates

An arXiv abstract reports a new attribution approach for networks whose raw coordinate outputs are mapped to angles, gaze directions or rotations by a fixed decoder, aiming to explain the final prediction rather than discarded components.

In plain words

Some networks predict directions or rotations by first outputting plain coordinates, which a fixed formula then converts into the final answer. The authors argue that explaining the raw coordinates can credit inputs the formula ignores. Their method instead explains the converted prediction, using standard attribution tools with decoder-aware targets.

What was announced

The abstract says conventional attribution applied to raw coordinates can reward inputs for changes the decoder discards, such as rescaling when only direction matters, so explanations may highlight inputs with no effect on the prediction. The authors instead attribute through the decoder differential, which for normalization decoders reduces to a scaled tangent projection, and they characterize invariance to output rescaling.

Reported experiments support the approach: on synthetic tasks with known relevance, ambient attribution assigned up to 72% of its mass to an irrelevant region; on MPIIGaze validation data, tangent Grad-CAM produced 44% greater prediction displacement under 10% pixel deletion; and on PASCAL3D+ with a 6D rotation decoder, the full differential improved mean deletion displacement by 14.1 degrees over an ambient coordinate sum and 13.8 degrees over a decoder-invariant scalar. Results were mixed for a two-polarity tangent LayerCAM variant, which beat ambient-sum LayerCAM on circular regression but showed no detected difference on quaternion pose. The authors note the method reuses existing attribution techniques with decoder-aware targets, though some variants need changes to attribution aggregation, and that an invariant attribution can still miss inputs driving the prediction, so invariance and deletion performance should each be checked.

What it does not show

  • Evidence is limited to the paper's abstract, so experimental details, full results and peer review status are not available.
  • The reported gains come from the authors' own experiments on specific datasets and have not been independently confirmed.
  • Mixed results, including no detected difference for one LayerCAM variant on quaternion pose, mean the method's advantage is not uniform across tasks.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.