Expert Level Detection of Pathologies From Unannotated Chess X-Ray Images

In check zero, we were able to look at pairs of images and reports. We're just going to learn to identify, does this report match this image? And you say, if the model says, yes, this report matches this image and it's right, we credit it. And otherwise, we penalize it. It is built on two works that inspired this idea. One of them is OpenAI's clip model. The other one is Stanford's convert model which was actually an original inspiration for clip. But this is an elegant idea that we were happy to find actually came to the same level of performance as we had shown in the early work.