StabiFreeze: Early Stopping for Training Binary Neural Networks via Internal Dynamics Stabilization

Binary neural networks (BNNs) offer significant efficiency in memory and computation during inference, making them promising for on-device applications. However, training BNNs remains computationally intensive, which limits their primary use to inference rather than ondevice training. This paper proposes StabiFreeze, a layerwise freezing framework that detects stabilization in convolution and batch normalization layers using a stabilitybased criterion. On CIFAR-100, StabiFreeze achieves up to $67.63 \%$ reduction in total training computation, completing training with $56.75 \%$ fewer epochs, while incurring accuracy degradation of only $0.30 \%$ and $0.54 \%$ for BinaryNet and XNOR-Net, respectively. Evaluation results on Raspberry Pi 4 and Jetson Nano show that StabiFreeze reduces training time by $80.46 \%$ and $62.56 \%$, with a $0.20 \%$ decrease and $0.07 \%$ increase in accuracy, respectively. These results demonstrate that StabiFreeze enables efficient and practical BNN training in resource-constrained edge environments.