14 DAY TRIAL // Synthetic neural networks could become the next big thing in music recognition and classification, a group of students from the University of Hong Kong believes. They say that the instruments, which were originally developed for applications in image recognition, are very well suited for actually seeing the sounds in a tune. By developing categories, and criteria to determine which sound goes in which category, researchers could conceivably develop a new music classification system that would divide songs into genres automatically, Technology Review reports.
Neural networks are groups of artificial neuron nodes, which can be programmed to handle a wide variety of tasks. The Hong Kong team managed to “train” such a network in recognizing songs, by subjecting it to a database of tunes spanning about 10 genres. After the neurons learned to extract traits such as tempo and harmony from a melody, the researchers fed them other songs, for classification. According to the study results, the synthetic neurons were capable of identifying the correct genre for the new tunes with an 87 percent accuracy rate. This is outstanding in the field of computer sciences, where music classification is considered to be a very tough nut to crack.
The convolutional neural networks used in this research were derived from knowledge of a cat's visual cortex, the researchers say, adding that this brain region is not very different in cats and humans. Oddly enough, past studies carried out on ferrets have determined that rewiring of their brains, not unlike what was recently done in the neural networks, can make the animal capable of distinguishing images using its auditory cortex. This type of research is bound to attract more and more interest in the years to come, due to the large number of potential applications its results may have in the real world.
The students admit that there's still a lot of work to be done on their network. They say that it fell short of reaching its goals when it was exposed to random songs. This means that it cannot yet be released into the “wild,” such as for instance on dedicated Internet websites. Its developers believe that this happens because the original library of songs on which it was trained was to narrow in focus. Exposing the networks to more songs, of varied genres, may help fix the problem.