Neural Network Coding Layer (NNCL): Enhancing Deep Learning Robustness Against Feature Erasure

This paper proposes a novel Neural Network Coding Layer (NNCL) that applies network coding theory to provide structured redundancy and enable reconstruction of lost features, thereby mitigating information loss problems in deep learning models. Unlike existing skip connections or attention mechanisms, NNCL embeds reconstructible redundancy into intermediate features through a learnable coding process, providing an explicit algebraic restoration mechanism for the original features. This design adapts the core principle of network coding—encoding information recoverably across a channel—for application within the computational graph of a neural network. Extensive experiments on the CIFAR-10 and CIFAR-100 datasets validate NNCL’s effectiveness. Even under a more conventional feature erasure rate of 20%, NNCL consistently improved classification accuracy by up to 8.3 percentage points. The model’s robustness becomes even more pronounced under extreme conditions. In a deliberate stress test involving 60% feature erasure—a scenario where baseline model performance collapses—NNCL dramatically boosted accuracy by up to 40.3 percentage points (e.g., from 19.8% to 60.1% on CIFAR-100). The proposed layer is designed to be modular and has been successfully integrated into various modern architectures, including ResNet and EfficientNet, and Vision Transformer, proving its broad applicability.