Hacker News Reader

I've been working on my generative adversarial network. It's a fascinating type of ML that can learn from both supervised and unsupervised data.

In essence, given data it will create a generator of that data. Right now my datasets are mostly images. I have tried on audio data and that does start to work - then diverges.

I'm still trying to find a configuration that works really well:

https://github.com/255bits/hypergan *

* needs modern GPU to run

View on HN

This is really interesting. How much data does it require to build a generative model with say 60% accuracy for story type data?

GANs 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

Reply on news.ycombinator.com