Evolution of the music industry artificial intelligence will automatically choose Love Song for users

November 12 News, according to foreign media reports, perhaps you are still listening to what songs to worry about, but in the near future the Internet music industry will adopt a more advanced way to solve this problem for you. Google, Baidu and Spotify have not yet fully displayed the full picture of the method, but they have struggled to provide a better playlist of music to users using an AI system called deep learning.

"Deep learning" is a training system branch of artificial intelligence, called "Artificial neural network". At present, all these companies have hired "deep learning" experts. Companies, including Google and Baidu, have used "depth learning" tools for a variety of purposes-advertising, speech recognition, image recognition, and even datacenter optimization. A start-up company even intends to use "depth learning" to identify patterns of medical images.

Now, these companies are turning to the music industry. The neural network based on music streaming media service can recognize the chord pattern of music without the guidance of musicians. Then you recommend songs, albums, or artists that match your preferences. Putting these complex systems into practical use is not a matter of overnight. But once the technology matures, "deep learning" may make it impossible for users to give up music streaming services in the future.

The origin of the music streaming media neural network

People began to pay attention to music streaming media neural network originated from the University of Ghent in Belgium last year's academic papers. The university's electronic and Information Systems Division published a paper entitled "Survival Calculations". This paper describes the method of selecting song attributes using convolution neural network (cellular Neural Network). Instead of using neural networks to observe the characteristics of images, as engineers did before.

The paper found that their approach could "produce reasonable advice". More importantly, their experiments show that the system is "significantly superior to the traditional approach". Microsoft researchers recently even cited the paper as an overview of the "Deep learning" field. The paper also attracted the attention of Spotify. "They invited us to go to Spotify's office to talk about the relevant content, and I think our paper is very helpful," said Dieleman, author of the paper. ”

The industry is widely accepted

"Deep learning" first emerged from Spotify's system. At present, Spotify uses more traditional data analysis methods to analyze the text content of specific music on the Internet and to analyze the song itself. Acoustic analysis is based on certain characteristics of the song, such as Rhythm, volume, and key. The system needs a large number of specific areas of information input. But the "deep learning" method created by the Belgians is completely different.

"Depth learning" analyses the sound waveform and assumes that we don't know the content of the song. Then the machine can automatically analyze all the results. It is a very versatile model and has great potential. The system did not sample Spotify's data, but provided playlists based on the similarity of the songs. The system is not the perfect substitute for the song-selection method Spotify uses. But the company believes it is something that deserves further study.

At the same time, "deep learning" has been used in a variety of Google, and its employees are certainly investigating the theory of music streaming media applications May. "Exciting ' deep learning ' represents a complete revolution, an absolute revolution," says Doug Ecques, a scientist who specializes in music technology. ”

But the trouble is that "deep learning" may do a good job of detecting the similarity of songs, but it is not easy to optimize the results of a selection that could mean testing multiple data. So "deep learning" may not be a simple replacement for music analysis software. It may be another tool that may not only be used to identify playlists, it has more potential than that. "What I see is that ' deep learning ' enables us to better understand music, so that we can really better understand what music is," says Doug Ecques. Google can use it to build better products, a better streaming media service. ”

The possibilities of the future

Most importantly, "deep learning" may help people discover new music. The music may not be popular at the moment, but users might like it once they hear it. But fundamentally, the future remains a challenge for the analytical systems that introduce new music to listeners. Many streaming media playback services have accumulated rich music data. Unfortunately, there is always new music, so streaming music service still needs a general analysis method.

In addition, "depth learning" is not just for music lovers. For artists, "deep learning" can also help them. Not based on sales reputation, but only the analysis of the music itself, can better help those independent musicians and new singers in time to be found by the audience.

(Responsible editor: Mengyishan)