Megadose The AI curve, compressed.

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

· ArXiv · AI/CL/LG ·
Standard attribution methods can credit inputs for changes that never reach a model’s final prediction.

For angle, gaze and rotation models, a fixed decoder can discard changes in raw outputs, leaving explanations that highlight irrelevant inputs. The paper traces attribution through that decoder; on synthetic tasks, conventional attribution assigned up to 72% of its mass to an irrelevant region. On MPIIGaze, tangent Grad-CAM produced 44% greater prediction displacement after deleting 10% of pixels. Results varied across methods and tasks, and the authors caution that decoder invariance alone does not guarantee an explanation captures what drives predictions.

ArXiv · AI/CL/LG's note

score 4

Categories: Research