CALM: Calibrated adaptive learning via mutual-ensemble fusion
Gading Aditya Perdana, Muhammad Alif Ghazali, Irene Anindaputri Iswanto, Setiawan Joddy · Procedia Computer Science · 2025
We introduce CALM (Calibrated Adaptive Learning via Mutual-Ensemble Fusion), a multi-stage framework for vision models that are both accurate and reliably calibrated. Stage 1 employs an Adaptive Curriculum Protocol (ACP) on CIFAR-10 to manage cross-entropy, distillation/mutual learning, feature alignment, and calibration losses for foundational students ( S d , S m ). A Het- erogeneous Feature Integration (HFI) module facilitates knowledge transfer from diverse teacher architectures. Stage 2 trains a meta-student ( S meta ) on CIFAR-10 to fuse knowledge from S d and S m via a learned combiner, potentially guided by a per-sample Adaptive Knowledge Transfer Protocol (AKTP). An optional Stage 2.5 further refines S meta ’s calibration on CIFAR-10 using a targeted loss. Finally, Stage 3 benchmarks S meta against minimally adapted baseline ( S b ), S d , and S m students on STL-10 to assess generalization and calibration robustness under domain shift. Evaluations demonstrate CALM’s effectiveness, with the recalibrated S meta showing strong performance and significantly reduced Expected Calibration Error (ECE) on CIFAR-10, and all models ex- hibiting informative calibration behavior when transferred to STL-10. CALM offers a systematic approach to producing efficient, uncertainty-aware models.