Performance Benchmarking and Optimization Analysis of Typical Classification Models Based on Binarized Neural Networks

This study addresses the fundamental trade-off between model compression and accuracy retention in Binarized Neural Networks (BNNs) by investigating gradient discontinuity induced by weight binarization and feature degradation stemming from activation discretization. Through systematic comparison of AlexNet and Network-in-Network (NIN) architectures on CIFAR-10, we propose a dynamic binarization-restoration mechanism with gradient direction preservation, integrated with topology-adaptive learning rate scheduling and channel attention weighting to mitigate information loss in texture-rich regions. Experimental results demonstrate that NIN's densely connected structure achieves significantly improved accuracy-efficiency balance compared to AlexNet, showing enhanced noise robustness and reduced parameter redundancy while maintaining competitive performance. The proposed framework provides actionable architectural optimization guidelines for implementing BNNs across heterogeneous computing platforms.