Increamental learning for classification

Say I’ve already trained my model to recognise 20 different workers using resnet34 on my own personal PC (with gpu), and now I am exporting it to a company server (no gpu) for it to be used.

If a new worker joins the company, is there a way I can train only for the new worker on the server? Or do I have to retrain all 21 workers from scratch? Any comments will be most welcomed, Thanks!


The phenomenon you are referring to is called catastrophic forgetting, there have been a few papers, with associated codes in the past years.

There have been multiple method proposed like using a sample from original dataset along with the new data to retrain the network, using bayesian inference to avoid forgetfulness, etc. though I have not seen any used in realtime solutions so far.