SEU-tolerant Restricted Boltzmann Machine learning on DSP-based fault detection
Songlei Jian, Jingfei Jiang, Kai Lü, Yanping Zhang · 2014
Restricted Boltzmann Machine (RBM) is the main building block of many deep learning models which are now becoming one of the most important kind of algorithms in machine leaning community. With the growing demand for information processing in aerospace computer, RBM can be used as a promising block supporting intelligent applications such as space self-control, intelligent recognition, and target classification. As for aerospace computer, a common challenge is the Single Event Upset (SEU) effect, which would change the state of some information bits stochastically and lead to error results. Due to the time-consuming and data-intensive computation of aerospace computer, digital signal processing (DSP) is the most suitable for aerospace computation. In this paper, we first implement representative RBM learning algorithm in a power efficient DSP platform. Then, we explore the possible SEU effects in our hardware architecture and RBM learning algorithm. Further, we integrate three software fault detection techniques (i.e. duplication, increasing data diversity, shortening life cycle of variable) into RBM learning. In the experiment, we utilize the simulated fault injection technique to evaluate fault detective RBM learning. The evaluation shows that our fault detection designs effectively detect SEU-induced error during RBM learning in DSP with low computational complexity.