Researchers from Zurich’s Institute of Embedded Systems at the University of Applied Sciences Winterthur have addressed the issue of reliability and safety in AI models. This is especially relevant for systems with essential safety integrated functions (SIF), such as edge-AI devices. The team noted that while existing redundancy techniques are effective, they are often computationally costly. They proposed a new method called “redundant execution,” which merges reliable model execution with non-reliable execution in a bid to minimize computational expenses while ensuring optimal performance.
“Redundant execution” implements the concept of a hybrid convolutional neural network (CNN) to facilitate reliable neural network execution. This hybrid network combines redundant execution techniques with standard CNN architectures to ensure the dependable execution of crucial operations while simultaneously minimizing computational resources.
The proposed approach was tested and found feasible by using the recognition of a “Stop” traffic sign as an example. The research found that integrating dependability features into a neural network is possible, but it does call for optimization. The hybrid CNN model narrows down the necessary dependable execution to extents that a dependable model determines, thus saving both footprint and computational power.
In essence, this proposed method addresses the urgent requirement to ensure the safety and reliability of AI systems, particularly in safety-critical applications. The method partitions the CNN into reliable and unreliable executions and incorporates reliable datasets into the training process. The hybrid network proves robust and safe without increasing the computational overhead significantly. The method may be further extended to more complex neural network architectures and applications with additional optimization. Overall, the redundant execution approach significantly contributes to creating safe and dependable AI systems.
