Dim(N) Week 6 - Yuhwan Kim (08/08/24)
Interpreting Nature and Humanity through Artificial Neural Networks (with a focus on the specific example of 'Music Cognition')
[Paper]
Spontaneous emergence of rudimentary music detectors in deep neural networks. Kim, G., Kim, DK. & Jeong, H. Nat Commun 15, 148 (2024). https://doi.org/10.1038/s41467-023-44516-0
[What Can Be Gained?]
The potential to actively utilize open data such as Audioset in neuroscience research.
Examples of creative use of artificial neural networks for testing neuroscientific hypotheses.
[Summary]
It was demonstrated that when an artificial neural network was trained with sound data from nature excluding music data, a music cognition function spontaneously emerged. (Not only did the feature vectors clearly distinguish between music and non-music, but they also categorized music types in detail.)
It was proven that there are neuron/network units that selectively respond to music. (Similarly, the neural network trained with non-music data showed this characteristic.)
It was shown that suppressing these network units reduced the ability to generalize natural sounds. Based on this, the hypothesis is proposed that music-selective neurons form the basis for the generalization of natural sounds and may directly originate from the ability to process natural sounds.
Comments