Reflective-net: learning from explanations

Abstract We examine whether data generated by explanation techniques, which promote a process of self-reflection, can improve classifier performance. Our work is based on the idea that humans have ability to make quick, intuitive decisions as well reflect their own thinking and learn from explanations. To best our knowledge, this first time potential mimicking using explanations explainability methods has been explored. found combining with traditional labeled leads significant improvements in classification accuracy training efficiency across multiple image datasets convolutional neural network architectures. It worth noting during training, we not only used for correct or predicted class, but also other classes. This serves purposes, including allowing reflection outcomes enriching through augmentation.