An Unequal Error Protection Scheme for Neural Network Parameters

As artificial intelligence progresses, neural networks are widely used these days. For parameters of neural network models, bit errors occurred during transmission or storage can significantly impact the overall model performance. Given the vast number of parameters in neural networks, it requires numerous coding redundancy overhead to protect the model parameters. Consequently, identifying an effective error protection scheme to protect model parameters with less coding redundancy is crucial. To this end, we design an unequal error protection scheme to protect neural network parameters. We provide efficient protection for model parameters by designing importance-based bit grouping method according to bit error sensitivity. Then for bit groups, we develop an algorithm with reinforcement learning to determine the precise coding redundancy allocation automatically. In comparison to the existing work, our method can maintain the original model performance while using fewer coding redundancy bits.