[Klathmon]

I doubt it, that would be a good chunk of data for what I see as a fairly small feature.

More likely (this is a guess, nobody outside google really knows at this point), they will use federated machine learning to figure out that something is "a song", then perhaps clean up and isolate the actual "song" part of it and send that over to a google server for processing.

But again it was just announced, so nobody really knows how this works, where the data is or goes, and what tradeoffs were made.

[angus-prune]

They announced, in the talk, that this is all done locally on the phone. The phone will have a database of, iirc, 10,000 songs locally.

[edit] typos

[Klathmon]

Ah! Then I'm sure they have some way of making the data needed much smaller than I thought it could be!

[Anm]

They only need to store (and occasionally update) the kernel parameters for the trained deep neural net. Very small indeed.

[user-on1]

how small, what technology is that?

[Anm]

I don't do neural nets, but if I had to crudely estimate... 10,000 songs (the outputs) * 16 layers * 16 parameters per node * 4 for the bytes per float = 10MB + a DB of song/artist names.

I'm probably underestimating the parameters per node, and overestimating the size of the layers closer to the input. Further, it's more likely structured as an LSTM than a convolutional network, since sound is a streaming source.