Abstract: Music exists in almost every society, has universal acoustic features, and is processed by distinct neural circuits in humans even with no experience of musical training. However, it remains unclear how these innate characteristics emerge and what functions they serve. Here, using an artificial deep neural network that models the auditory information processing of the brain, we show that units tuned to music can spontaneously emerge by learning natural sound detection, even without learning music. The music-selective units encoded the temporal structure of music in multiple timescales, following the population-level response characteristics observed in the brain. We found that the process of generalization is critical for the emergence of music-selectivity and that music-selectivity can work as a functional basis for the generalization of natural sound, thereby elucidating its origin. These findings suggest that evolutionary adaptation to process natural sounds can provide an initial blueprint for our sense of music.
Introduction
In the human brain, specific neural populations in the auditory cortex respond selectively to music, regardless of whether a person has had formal musical training. Even infants show a preference for musical features like melody and tempo. This raises a fundamental question: Is music-selectivity a result of lifelong exposure to songs, or is it an innate "blueprint" hardwired into us? Because it is nearly impossible to find humans with zero exposure to music, this study uses artificial neural networks to test whether the machinery for music processing can emerge simply from the need to understand natural, non-musical sounds.
Methods and Datasets
The researchers built a Convolutional Neural Network (CNN) designed for audio event detection, featuring four convolutional layers followed by global average pooling to mimic the human auditory system.
The network was trained on AudioSet, a massive collection of 10-second YouTube clips. Crucially, the authors explicitly excluded all music-related categories from the training data, focusing on 527 categories like animal noises, speech, and environmental sounds. To analyze the resulting internal representations, they used a Music Selectivity Index (MSI) to quantify unit responses and t-SNE visualization to map high-dimensional neural representations into a 2D space.
Key Results
Despite never being trained on music, the network spontaneously developed "music-selective" units. These units showed response profiles remarkably similar to music-selective neurons found in human fMRI studies.
Key findings include:
- Distinct Representation: A linear classifier could distinguish music from non-music with high accuracy (mAP ~0.887), proving music features were encoded in the internal representations. t-SNE visualization showed clear clustering of musical genres (Jazz, Reggae, Soul) separate from non-musical sounds (Speech, Vehicles).
- Strong Selectivity: The top 12.5% of units (by MSI) showed an average of 2.07 times stronger response to music than to other sounds.
- Generalization and Function: Music-selectivity only emerged when the network was forced to generalize patterns rather than memorize labels. Furthermore, "silencing" these music-selective units significantly degraded the network's performance on general natural sound detection, suggesting that music-like features are fundamental building blocks for understanding the natural auditory world.
- Temporal Sensitivity: These artificial detectors were sensitive to the complex temporal structure of music across multiple timescales, mirroring biological observations.
Discussion
These findings suggest that our biological "sense of music" may arise as a functional byproduct of evolutionary adaptation to process the natural auditory environment. However, the study notes limitations: the feedforward CNN architecture lacks the recurrent feedback of the human brain, and the training data includes modern sounds (like trains) that differ from the ancestral environment. Future work may explore if similar spontaneous detectors emerge for other cultural universals, such as face perception or written language, through the lens of natural scene statistics.
Summary updated