David Kelly compares music classifiers to Clever Hans, the horse whose apparent arithmetic could be explained by cues from a questioner. He explains how frequency-based shortcuts can give a model a correct-looking label without the musical understanding a listener expects.
David Kelly decomposes audio into frequency components and uses a divide-and-conquer search for signals sufficient to preserve a classification. He distinguishes those from signals that are also necessary: removing them changes the model’s answer.
In an illustrative blues recording, David Kelly plays a minimal signal classified as blues and then removes frequencies so the remaining audio is classified as hip hop. He cautions that the example was chosen to make the difference clear and describes the challenge of interpreting accuracy as understanding.
Watch on YouTube
