[gurovich]

This resource has been invaluable to me: https://paperswithcode.com/

From the past examples you give it sounds like you were into computer vision. There’s been a ton of developments since then, and I think you’d really enjoy the applications of some of those classic convolutional and variational encoder techniques in combination with transformers. A state of the art multimodal non-autoregressive neural net model such as Google’s Muse is a nice paper to work up to, since it exposes a breadth of approaches.