This document discusses a novel test-cost-sensitive method for training convolutional neural networks (CNNs) that incorporates expert branches to determine the optimal depth of the network based on the difficulty of input instances, effectively balancing accuracy and computational cost. Experimental results illustrate that the proposed method achieves lower test-costs while maintaining competitive accuracy compared to traditional models. Additionally, it highlights advancements in facial expression detection using local feature extraction algorithms, emphasizing the superiority of local algorithms over holistic ones in recognition tasks.