5 ms·
If you can have humans sort the generated images into "good quality" and "bad quality", you can just keep iterating. Our subjective ratings is another score to
by stocknoob 4y ago
If you can have humans sort the generated images into "good quality" and "bad quality", you can just keep iterating. Our subjective ratings is another score to optimize for.
- cloogshicer 4y agoDoesn't the sample size for this have to be very large for it to make a difference? Genuine question.
- stocknoob 4y agoSure, but there are millions of people on the DALLE waitlist, who would happily rate the output for better performance / more credits. The famous ImageNet data set only has 1.2M images.
- spyder 4y agoWith semi-supervised learning a small amount of labeled data, can produce considerable improvement in accuracy: https://en.wikipedia.org/wiki/Semi-supervised_learning https://en.wikipedia.org/wiki/Semi-supervised_learning https://towardsdatascience.com/semi-supervised-learning-how-to-assign-labels-with-label-propagation-algorithm-9f1683f4d0eb https://towardsdatascience.com/semi-supervised-learning-how-...
- cloogshicer 4y agoThank you!
- cube2222 4y agoMoreover, the current Dalle UI already does that. When you run a phrase, you get four images. Those images will stay in your history, but the ones you like you will save with the "save" button, so that they're in your private collection. With this, you already have a great feedback system: saved - good, not saved - bad.
- rockemsockem 4y agoI've saved some of the worst images Dalle generated to be able to showcase just how bad it can be sometimes. And then other times the bad image is hilariously bad. They can probably build another layer on top of the feedback system though to filter that sort of thing out.
- nyhc99 4y agoI would guess your use-case is a statistical anomaly. If most of the images that are saved are saved by people who like them best, which is most likely the case, enough data will erase the problem.