Great post! It was fun to see 4 tracks in the (first convolutional layer) filter 242 set that I recognized from my own 'ambient' Spotify playlists, and pretty impressive at the topmost layer as well. Loved that approach of looking at a few tracks that represented maximally or average activated filters.
Curious if you think the low-level features learned from the vector_exp latent factors are different from, say, unsupervised learning with sparse autoencoders? For example, are there phonemes associated with Chinese pop or Spanish rap that are learned at a low level, that the network might not learn with "unlabeled" data?
I haven't tried it, so I can't say for sure, but my intuition is that learning to predict the latent factors is a much less 'complex' task than learning good features to reconstruct the input (i.e. the spectrograms), in terms of the required capacity of the model.
With a purely unsupervised approach, you are basically wasting capacity on modeling aspects of the data that are relevant for reconstructing the input, but not for solving the task at hand. For example, the model doesn't have to care about precise pitches and timing, because those are not relevant for recommendation (and latent factor prediction) anyway. That means no model capacity is wasted on these things. With a fairly complex task such as this one, I think that probably makes a big difference.
Curious if you think the low-level features learned from the vector_exp latent factors are different from, say, unsupervised learning with sparse autoencoders? For example, are there phonemes associated with Chinese pop or Spanish rap that are learned at a low level, that the network might not learn with "unlabeled" data?