How many names in a dataset are required for a RNN for name generation to produce original names? It’s possible to do so with just 20 names?
How many names in a dataset are required for a RNN for name generation to produce original names? It’s possible to do so with just 20 names?
Why are you trying to use an RNN instead of something like a Markov chain for that?
@e0qdk@reddthat.com I heard that RNNs for this task are of higher quality
I don’t have an answer to your original question – I think it would likely depend on exactly how you set up your network, how you encode your data and decode your output, what qualifies as “original” etc. I’m mostly familiar with RNNs from their relation to IIR filters (in audio) than however they are (or aren’t) used in ML these days… but the idea of trying to set something like that up to deliberately spit out names just makes me feel tired thinking about it.
If you’re looking to do this for procgen in gamedev or something, a Markov chain (state machine that follows transitions based on probability) is likely to be a lot easier to reason about.