6 ms·
This is really interesting. How much data does it require to build a generative model with say 60% accuracy for story type data?
by infinitone 10y ago
This is really interesting. How much data does it require to build a generative model with say 60% accuracy for story type data?
- whataretensors 10y agoGANs work on real-valued data. The problem with sequence data, like text, is that the error gradient can't flow through the generator. There's some really great work being done called "Sequence GAN" that turns sequence generation into a RL problem. I'm not sure how successful they are, but it seems potentially powerful. https://arxiv.org/abs/1609.05473 https://arxiv.org/abs/1609.05473