The research introduces a statistical framework for early time classification algorithms, which aim to label a stream of features without processing the full input stream. The framework formulates a calibrated stopping rule, controlling the accuracy gap between full and early-time classification. The early stopping mechanism reduces up to 94% of time steps used for classification while maintaining rigorous accuracy control. This method is particularly useful in scenarios requiring quick predictive inferences, such as reading comprehension tasks, real-time song identification, and computational tomography systems.
Publication date: 1 Feb 2024
Project Page: https://arxiv.org/abs/2402.00857
Paper: https://arxiv.org/pdf/2402.00857
