Improving CNN Runtime Robustness Against Soft Errors by Dropout Layer Optimization

Convolutional Neural Networks (CNNs) have shown exceptional effectiveness in complex and data-intensive domains such as image and video processing, conversational systems, and healthcare. Moreover, sectors like High-Performance Computing and safety-critical applications, including automotive, aerospace, and autonomous robotics, impose stringent requirements on energy efficiency, performance, and robustness. However, modern semiconductor technologies are increasingly vulnerable to faults, which can degrade CNN performance and potentially result in catastrophic failures. This work explores the impact of regularization techniques (dropout layer) in enhancing the inference robustness of CNN models against soft errors. We analyzed soft error impacts on five widely adopted CNN architectures, each trained with ten different dropout rates. Our experimental results reveal that optimizing the dropout rate during training can improve the in-field robustness of CNN models by up to 12% compared to baseline configurations under soft error conditions. Additionally, fine-tuning this architectural parameter can lead to accuracy improvements of up to 10%.