Use of Metamorphic Relations as Knowledge Carriers to Train Deep Neural Networks
This addresses the problem of ineffective DNN training for AI researchers, but it is incremental as it builds on existing metamorphic testing concepts and requires more work for broader impact.
The paper tackles the difficulty of training deep neural networks (DNNs) by introducing an approach using metamorphic relations (MRs) as knowledge carriers, and a preliminary experiment shows that a DNN trained with MRs delivers better performance than one trained without MRs.
Training multiple-layered deep neural networks (DNNs) is difficult. The standard practice of using a large number of samples for training often does not improve the performance of a DNN to a satisfactory level. Thus, a systematic training approach is needed. To address this need, we introduce an innovative approach of using metamorphic relations (MRs) as "knowledge carriers" to train DNNs. Based on the concept of metamorphic testing and MRs (which play the role of a test oracle in software testing), we make use of the notion of metamorphic group of inputs as concrete instances of MRs (which are abstractions of knowledge) to train a DNN in a systematic and effective manner. To verify the viability of our training approach, we have conducted a preliminary experiment to compare the performance of two DNNs: one trained with MRs and the other trained without MRs. We found that the DNN trained with MRs has delivered a better performance, thereby confirming that our approach of using MRs as knowledge carriers to train DNNs is promising. More work and studies, however, are needed to solidify and leverage this approach to generate widespread impact on effective DNN training.