[dangravell]

At the back end of the 90s I did some work with self-organizing maps - collapsing multi dimensional space into a 2D map. I think it's interesting as a data visualisation approach. This looks similar but I'm totally out of date with where SOM went.

When you say 128d space, what are the data that are represented, exactly?

[deppep]

Hello! I find these fascinating too. The 128 dimensions I mentioned are learned by the NN, they are latent space and their actual meaning is ultimately unknown. The number itself is somewhat arbitrary as it was selected optimizing memory usage vs data coverage so that the track embeddings could sit on an A100.