Reliable Classification with Ensemble Convolutional Neural Networks

Convolutional Neural Networks (CNNs) are widely used in computer vision and natural language processing. Due to large computational requirements, implementation of CNNs on FPGAs becomes an popular option. As CNNs being used in safety critical applications, reliability become a priority. This poses challenges as FPGAs are prone to suffer soft errors. Traditional fault tolerant techniques based on modular redundancy introduce a large overhead, which may not be acceptable for many resources-limited embedded system. This paper explores the use of an ensemble of CNNs to build reliable classifiers. The idea is to combine several "weak" classifiers to obtain a "strong" one, so that the classifier can still work reliably if one of its members fails. Differently from traditional ensemble learning that looks for the classifiers to complement each other, in our case similarity is also important to achieve fault tolerance. To evaluate the potential of using ensembles to implement fault tolerant CNNs, an initial study is done on ResNets. The results show that, relative to a single deep ResNet, an ensemble of shallow ResNets could provide similar classification results while providing an effective protection against errors with limited overhead.