This is likely to be something we try as part of our PRISM scenario, so we should start thinking about the scientific and technical requirements for a pipeline to do this. Things to consider are:
how does the pipeline differ from training from scratch (cold start) vs retraining (warm start)?
How might we change hyperparameter in such a scenario (e.g. learning rate) so it doesn't overfit/underfit either the new data or the old data?
How can we select data for retraining ?
How do we compare performance and measure gains from the retraining, if there are any? (probably several models available in parallel in. "parallel suite" sort of paradigm.
This is likely to be something we try as part of our PRISM scenario, so we should start thinking about the scientific and technical requirements for a pipeline to do this. Things to consider are: