CPATTA: Conformal Supervision Allocation For Active Test-Time Adaptation
The episode discusses CPATTA: Conformal Supervision Allocation For Active Test-Time Adaptation, a paper by Shi, Lyu, and Peng. The hosts explore how CPATTA uses smoothed conformal scores and top-K certainty measures to guide which samples get labeled during deployment. The work improves existing methods by using online feedback loops and domain shift detection to make adaptation more robust.