Inference-Time Decision Calibration for Temporal Classification
This paper introduces a novel approach to temporal classification by decomposing errors into representation failures and decision-making issues. The proposed method involves freezing a trained classifier and adding two inference-time interventions: a multi-scale residual branch for auxiliary logits and a branch-aware calibrator to recombine evidence. Experiments on various datasets like FI-2010 and PTB-XL show that these interventions yield significant gains, particularly in noisy or representation-limited scenarios, suggesting that temporal classification benefits from improved evidence calibration alongside representation learning. AI