Abstract
This is about a phenomenon we found while we were applying LSpec to MNIST dataset for quick experiments. We mixed tampered digit images (each one is generated by mixing two images of two different digits, and assigning random label) in the training dataset, and tried to make LSpec to highlight these tampered inputs. The LSpec was successful in highlighting these "bad" records. But we also found another unexpected phenomenon, which is improved learning speed. We are going to investigate the reason of this phenomenon, trying to explain with existing theories, and further move on to apply to ultrasound dataset which is the main focus of our other research topics.