I've looked at this before to some extent. The problem seems to me to be that (at least for training) adapting a model in 'real-time' in response to new input just seems too hard to do in practice. In reality when training a model there are usually lots of tweaks, fine-tuning and iteration needed that require human intervention, at which point you don't really need real-time anymore. Perhaps there is a case for using such approaches for inference although it's not something I've thought about too much.