Self-Healing Neural Networks for Resilient AI Systems

Robustness and adaptability are becoming imperative in modern AI systems, especially in dynamically-changing or failure-prone environments. Standard neural networks do not conceive their own mechanisms to respond to such internal faults, resulting in performance degradation or even a complete breakdown of the system. Proposed here is an innovative Self-Healing Neural Network (SHNN) architecture capable of self-monitoring, localizing, and repairing faults in its computation graph autonomously and externally cognizant of the surrounding world. SHNN incorporates the lightweight fault-monitoring mechanism and an adaptable retraining module to reconstruct the damaged nodes or layers by real-time performance feedbacks. Through experimental evaluations on benchmark datasets, we show that the model robustness of SHNNs can achieve up to 95% accuracy when up to 30% of its internal components are corrupted. This strategy leads to notably improved robustness of the whole system while incurring negligible computational overheads which happen to be an essential requirement for on-device deployment on self-sustainable AI such as in-persistent computing, autonomous cars, and other mission-critical applications. The suggested methodology encourages lifelong learning and recovering from faults and fits nicely into the vision of sustainable and self-managing AI systems. The findings further confirm the efficacy of SHNNs, as well as the potential for a new breed of resilient AI infrastructures that can cope well with perturbations in performance.