Lightweight Test-Time Adaptation Framework for Binary Neural Networks Using BN Recalibration and Entropy Minimization

This paper proposes a method to address the performance degradation problem in Binary Neural Networks (BNNs) caused by distribution mismatch between the training and inference stages [1], [2]. The proposed approach consists of simple test-time adaptation techniques that either directly recalibrate the Batch Normalization (BN) statistics [3] or minimize the prediction entropy to reduce model uncertainty [4]. Two modes are compared and analyzed: BN_ONLY, which updates only the BN statistics, and TENT (Test-time Entropy Minimization), which fine-tunes the BN scale and shift parameters while minimizing entropy [4]. Unlabeled data streams are continuously fed into the model, and performance is periodically evaluated using a validation dataset. Experimental results show that even simple BN recalibration improves accuracy by approximately 2.1%, while the TENT mode enhances prediction stability and reduces uncertainty [4],[5]. The proposed method provides a lightweight framework that enables adaptation to real-world environments without additional training, demonstrating its potential to improve domain generalization performance [6] in BNNs.