I had the pleasure of conducting research into self-supervised learning (SSL) for computer vision.
What stood out to me was the simplicity of the SSL algorithms combined with the astonishing performance of the self-supervisedly trained models after supervised fine-tuning.
Also the fact that SSL works across tasks and domains, e.g., text generation, image generation, semantic segmentation...
I am excited about continual reinforcement learning (RL). When I first learned about RL, I thought that it was too general for its own good. And yet, continual learning lies outside of the scope of current RL fundamental research. It's an exciting time because very little is understood about how reinforcement learning methods work with neural networks on simple problems. Yet, many interesting problems require not only RL but continual learning (either because the environment is changing in some unknown way or the environment may include interaction with a human). We are still at the very early stages, but I expect there to be synergy with current developments like LLMs.
Do you happen to have a good source to read up on continual RL that you can recommend? I am not familiar with this use case for RL.