ER-NeRV: Training Error-Resilient Neural Representations for Videos

Neural representations for videos (NeRV) are emerging as a learning-based alternative to traditional video codecs, enabling efficient and low-complexity playback across consumer and edge devices. These neural decoders increasingly operate on energy-optimized hardware where memory faults can degrade perceptual quality. In this study, we investigate the robustness of NeRV under simulated hardware-induced faults. We introduce ER-NeRV (Error-Resilient Neural Representations for Videos): a robust and efficient neural video codec to improve NeRV's resilience to such errors while maintaining its efficiency. By exposing the models to controlled bit-flip noise, ER-NeRV improves robustness to numerical faults possible in low-power edge execution. Compared to NeRV, we demonstrate that ER-NeRV can achieve a higher peak signal-to-noise ratio (PSNR) and multi-scale structural similarity index (MS-SSIM) under increasing bit error rates across variable bit-width regimes, while maintaining visual fidelity with negligible overhead.