Architectures to Ensure the Functional Safety of Neural Network based Systems
Published : 27 February 2019
Neural networks are increasingly used in mission critical systems such as those used for image recognition in autonomous vehicles. These systems must comply with standards for functional safety, therefore it is essential to ensure they operate correctly in the presence of certain types of faults and that they can detect those faults which could result in dangerous situations.
The same formal neural network can be implemented on different hardware platforms (CPUs, FPGAs, etc.), depending on the required performance. In some cases, implementations based on spike coding and neurons can result in significant power savings.
It is well understood how to analyze and improve the reliability of classical digital circuits (micro-controllers, RAMs, etc.), however, these approaches are not directly applicable to neural networks, especially those using spike coding and analog neurons.
The goal of this PhD thesis is to develop new approaches to improve the fault tolerance of spiking neural networks. As the first part of the thesis, new fault models and quantitative metrics to measure the correct operation of the system will be developed. Test cases using both classic coding and spiking networks will be prepared, to provide a reference for the studies. These will include cases using both off-line learning and unsupervised learning. Then the candidate will look for new techniques for detecting and managing faults in order to make the full system more robust. One avenue will be techniques for testing the system while it is operational (on-line test). Another research direction consists of studying how the architecture of the formal network and training data can be adapted to improve fault tolerance.