Exploring Learning with Deep Heterogeneous Descriptor-based Sampling

The training of deep learning models becomes more complex with the advancement of technology. Specifically, advanced computation hardware has a significant effect on the speed at which models can be trained. Improvements in training procedures are more likely to reduce time and capital costs. As a rule, samples have unequal importance in a dataset. Early model training will be accelerated with the help of informative samples. In order to maximize diversity of features, we develop DHDS (Deep Heterogeneous Descriptor-based Sampling), a novel method based on instance-level diversity sampling, which clusters examples according to deep similarity metrics. Furthermore, we perform extensive experiments, and the result demonstrates the DHDS approach accelerates network training and improves accuracy during the initial training stage.