Zero-Gate Residual Networks

Deep residual networks (ResNet) have become the de facto backbone for modern deep learning. However, vanilla ResNet v1 and pre‑activation ResNet v2 suffer from inherent structural drawbacks, including compromised identity shortcut paths, empirical dependence on Batch Normalization (BN), and suboptimal optimization landscapes. In this work, we propose a novel Zero‑Gate Residual Network (ZGRN) guided by gradient dynamics analysis. We first construct a pure identity main path to support distortion‑free signal propagation. We then introduce channel‑wise zero‑initialized learnable gating to replace BN and fixed handcrafted scaling factors, enabling progressive residual learning and adaptive feature calibration. Furthermore, we equip the residual branch with lightweight bilinear feature fusion to strengthen non‑linear representation capacity with controlled computational overhead. We provide rigorous mathematical analysis for the core properties of our design, including exact identity mapping at initialization and structural safeguards against gradient attenuation. Extensive experiments on CIFAR‑10, CIFAR‑100 and ImageNet‑1K demonstrate that ZGRN consistently outperforms vanilla ResNet variants in classification accuracy, training convergence speed and practical inference efficiency. Comprehensive ablation studies quantitatively validate the individual contribution of each proposed component. While our empirical validation is strictly limited to CNN structures in this work, the core design principle---pure identity path plus progressive residual learning---is conceptually generalizable to various deep stacked architectures including Transformers and diffusion models.