Music, a cultural universal, presents a fascinating enigma: its near-ubiquitous presence across societies and its processing by dedicated neural circuits in humans, even without musical training. This raises fundamental questions about its origins and evolutionary purpose. While music-selective neural populations exist in the brain's non-primary auditory cortex, responding specifically to music and not to speech or other environmental sounds, it remains unclear how these specialized circuits develop. Studies suggest that music selectivity develops spontaneously, even in individuals with minimal musical training, and even infants show an ability to perceive various acoustic features of music. This raises the possibility that innate predispositions, rather than solely experience-dependent learning, play a critical role. The role of lifelong passive exposure to music is debated, given that basic music processing mechanisms are observed even in populations with limited exposure to complex musical structures. This paper uses artificial deep neural networks (DNNs) to explore a hypothesis: music selectivity might arise as a by-product of the brain's adaptation to process natural sounds, suggesting that statistical patterns in natural sounds may constrain the innate basis for music perception. By training a DNN on natural sound detection, the researchers aim to demonstrate that music can be distinctly represented even in the absence of music in the training data.

Previous research highlights the universality of music across cultures, emphasizing common elements found worldwide. The perception and production of music are intrinsically linked to the brain's ability to process musical elements. Studies using neuroimaging techniques have revealed music-selective neural populations in the non-primary auditory cortex, demonstrating a specialized neural circuitry for music processing. These studies also suggest a degree of innate capacity for music perception, as music selectivity is found even in individuals without formal musical training. Furthermore, infants demonstrate an ability to perceive multiple acoustic features of music, indicating that some aspects of music processing may be pre-wired. While lifelong passive exposure to music may contribute, the existence of music processing mechanisms in cultures with limited exposure to complex musical structures suggests that innate predispositions are also at play. Recent DNN models have successfully replicated aspects of brain function, suggesting their potential to illuminate the principles underlying sensory processing and even higher-level cognitive functions. These models demonstrate that brain-like functional encoding of sensory inputs can emerge as a by-product of optimization to process natural stimuli. This suggests a potential pathway to understand the emergence of music-selectivity.

The researchers used the AudioSet dataset, comprising 10-second real-world audio excerpts from YouTube videos labeled with 527 categories of natural sounds. A conventional convolutional neural network (CNN) was designed, mirroring architectures used successfully in audio event detection and modeling human auditory cortex information processing. The network was trained on two conditions: one included music-related categories, and the other excluded them. The network's performance on audio event detection was evaluated. Following training, t-distributed stochastic neighbor embedding (t-SNE) was used to visualize the distribution of feature vectors extracted from the average pooling layer of the network. This aimed to determine if music data clustered distinctly from other sounds, both when music was and wasn't present in the training data. Linear models (PCA and GBFB) were used as a comparison to the CNN to assess if non-linear feature extraction was necessary for music selectivity. The music-selectivity index (MSI) of each unit was calculated to identify music-selective units. The response of these units to various stimuli was then analyzed. To investigate the encoding of temporal structure, a 'sound quilting' method was employed, reordering segments of audio to disrupt the long-range temporal structure while preserving short-range properties. Finally, an ablation study examined the impact of silencing music-selective units on the network's performance to assess their functional role. The effects of removing speech from the training data were also analyzed.

The key findings of this study are: 1. **Emergence of Music Selectivity without Music Training:** The DNN trained without music still showed a distinct representation of music, suggesting that music selectivity can emerge as a by-product of learning natural sound processing. This was not observed using linear feature extraction methods such as PCA or GBFB, demonstrating the necessity of non-linear feature extraction. 2. **Music-Selective Units:** Units within the network exhibited music-selective responses, showing significantly stronger activation to music than to other sounds. This music selectivity was robust to variations in sound amplitude. 3. **Temporal Structure Encoding:** Music-selective units encoded the temporal structure of music across multiple timescales, consistent with observations in the human brain. The response of these units was sensitive to segment size in sound quilts, reducing as segment size decreased, demonstrating sensitivity to long-range temporal structure. 4. **Generalization's Importance:** Music selectivity is shown to be closely tied to the network's ability to generalize across natural sounds. A network trained to memorize randomized labels, and therefore unable to generalize, lacked the characteristic temporal structure encoding. 5. **Functional Role of Music-Selective Units:** Ablation of music-selective units significantly impaired the network's ability to detect natural sounds, highlighting their functional importance beyond music processing. The network’s ability to perform the task was decreased more when music selective units were removed than when a similar number of units with lower MSI were removed, highlighting the importance of these units for generalization. 6. **Role of Speech:** While music selectivity emerged even without speech in the training data, the inclusion of speech enhanced the encoding of long-range temporal structures in music.

This study provides compelling evidence that music selectivity can spontaneously emerge in DNNs trained on natural sounds, without explicit music training. The findings support the hypothesis that the brain's adaptation to natural sounds may provide a foundational basis for music processing. The observed similarities between the DNN's music-selective units and those in the human brain suggest a potential shared underlying mechanism. The importance of generalization in the emergence of music selectivity highlights the interconnectedness between music and the broader processing of natural sounds. The robust encoding of temporal structure and the functional role of music-selective units in natural sound detection further strengthen this connection. The study's limitations, primarily related to the DNN model's simplicity compared to the biological brain's complexity, are acknowledged, but do not undermine the significance of the findings.

This research demonstrates that music selectivity can spontaneously emerge in DNNs trained on real-world natural sounds, even in the absence of music in the training data. The findings highlight the role of generalization in the development of music selectivity, suggesting that adaptation to natural sound processing may provide a crucial initial blueprint for music perception. Future research could investigate how experience-dependent factors shape and refine this innate capacity for music processing, and examine the phylogenetic distribution of music-processing abilities across species.

The study's limitations include the use of a simplified DNN model that doesn't fully capture the intricacies of the biological brain. The feedforward architecture of the CNN does not reflect the complex intracortical and top-down connections in the brain. Furthermore, the backpropagation learning mechanism in DNNs differs from the learning mechanisms in the brain. The specific dataset used in the study may also influence the results, and further research is needed to confirm the findings across a range of datasets and experimental conditions.