Reinforcement Learning-Based Approach to Protect Neural Networks from Soft Errors

Strategic Insights -

A new research report from Stockholm, Sweden, by KTH Royal Institute of Technology, has made significant advancements in sensor research, specifically focusing on protecting neural networks from soft errors. The study, funded by Kth Royal Institute of Technology with The Industrial Research Project Adinsos, employed a Reinforcement-Learning-based approach to address the vulnerability of Deep Neural Networks to soft errors. The research team, led by Peng Su, demonstrated the effectiveness of their method by achieving a performance gain of at least 10% to 15% over baseline methods.

Key Takeaways:

  • The research team proposed a Reinforcement-Learning-based approach to protect neural networks from soft errors by identifying and addressing vulnerable bits.
  • The approach consists of three key steps: layer-wise resiliency analysis, bit mask generation, and bit mask deployment.
  • The method was tested and validated on several existing neural networks, including Hamming code and Most Significant Bits protection schemes.
  • The results indicate that the proposed approach exhibits a significant improvement in performance compared to baseline methods.
  • The study highlights the importance of ensuring the robustness of neural networks in mission-critical applications.
  • Peng Su and his team demonstrated the effectiveness of their method in protecting neural networks from soft errors.

Statistics:

  • The proposed method achieved a performance gain of at least 10% to 15% over baseline methods.
  • The study analyzed layer-wise resiliency of Deep Neural Networks by a fault injection simulation.
  • The method generated layer-wise bit masks by a Reinforcement-Learning-based agent to reveal vulnerable bits.
  • The performance of the proposed approach was compared with Hamming code and Most Significant Bits protection schemes.

Sources:

  • Applying Reinforcement Learning to Protect Deep Neural Networks from Soft Errors. Sensors, 2025,25(13):4196. (Sensors - http://www.mdpi.com/journal/sensors)
  • NewsRx. Studies from KTH Royal Institute of Technology Further Understanding of Sensor Research (Applying Reinforcement Learning to Protect Deep Neural Networks from Soft Errors). Journal of Engineering. July 28, 2025; p 3358