I was testing multiple MIR tasks with my pre-trained features from compact_cnn. I thought I was.
It turned out that I didn’t even load the weights file. There is simply no such code. (You’re supposed to laugh at me now. out loud.)
The featured (thumbnail) image is what I ended up doing. Please scroll up and take a look. But I almost never realised it and even wrote quite a lot of ‘discussion’ on the top of the result. (Yeah, laugh at me. LOUDER) Anyway, it means my system was a deep convolutional extreme learning machine.
I couldn’t realised it earlier because the results are quite good. Let’s see how the feature worked with SVM classifier.
Not bad, huh?
Finally, Kyunghyun noted out that it has been known since 1965 by Cover’s theorem.
A complex pattern-classification problem, cast in a high-dimensional space nonlinearly, is more likely to be linearly separable than in a low-dimensional space, provided that the space is not densely populated.
— Cover, T.M., Geometrical and Statistical properties of systems of linear inequalities with applications in pattern recognition, 1965
Still quite interesting. Oh, and the non-ELM results, what I thought I was doing, will be posted to arXiv soon. See you there then.
PS. Saxe et al 2011, a NIPS paper about it. Deeplearningbook also mentions in Convnet chapter that “Random filters often work surprisingly well in convolutional networks”.
Did you actually lose performance when you increase the fft filter length and hop length while extracting the spectogram? For instance, I guess you have used n_fft = 512 and hop_length = 256. What happens when you have n_fft > 4000 and hop_length > 2000?
LikeLike
I haven’t tried, but I guess it will if it varies a lot – like you said. One reason to develop Kapre was to test such stuff, will do some of them quite soon.
LikeLike
y
LikeLike
I think music preference is identified by attention spikes more than 100ms. proabably network can perform better.
LikeLiked by 1 person