Applying Reinforcement Learning to Protect Deep Neural Networks from Soft Errors.

Area of Science:

• Artificial Intelligence
• Computer Science
• Electrical Engineering

Background:

• Deep Neural Networks (DNNs) are crucial for sensor-based systems but susceptible to soft errors, threatening system safety.
• Conventional fault tolerance methods face scalability challenges with complex DNN architectures.
• Ensuring DNN robustness against errors is vital for reliable AI applications.

Purpose of the Study:

• To develop an effective and scalable method for protecting DNNs against soft errors.
• To identify and mitigate vulnerable bits within DNNs using a novel approach.
• To enhance the safety and reliability of AI-powered sensor systems.

Main Methods:

• A Reinforcement Learning (RL)-based agent was developed to identify vulnerable bits in DNNs.
• Fault injection simulations were used to analyze layer-wise network resiliency.
• Transfer learning was employed for efficient synthesis and deployment of bit masks.

Main Results:

• The proposed RL-based method demonstrated significant performance gains (10-15%) over baseline techniques.
• Vulnerable bits were dynamically and efficiently protected, enhancing network robustness.
• The approach proved effective in protecting selected neural networks against soft errors.

Conclusions:

• The RL-based approach offers a superior solution for protecting DNNs from soft errors compared to traditional methods.
• This technique enhances the safety and reliability of AI systems by mitigating soft error impacts.
• The method provides a scalable and efficient way to improve DNN resilience.