Spontaneous emergence of rudimentary music detectors in deep neural networks

This study investigates the spontaneous emergence of music-selective units in deep neural networks (DNNs) trained on natural sounds. Using a DNN model of auditory information processing, the researchers demonstrate that units tuned to music emerge without explicit music training. These units encode the temporal structure of music across multiple timescales, mirroring brain responses. Generalization ability is identified as crucial for music selectivity's emergence, suggesting that adaptation to natural sound processing may be a foundational blueprint for our sense of music.

Gwangsu Kim, Dong-Kyum Kim, Hawoong Jeong