Ideas for multistage NN training.
OML paper derives some code from MAML:
The target for me is to create a multistage network that is suitable for online learning from a stream of natural data, e.g. video or timeseries. That setup is very similar to the online few shot learning they explore in ANML or OML. I think it makes sense to understand the background paper (MAML) first. So I would start with reading and understanding it’s code and reproducing the results.
However there is a paper MAML++ pdf code that is a better version of MAML. It’s much easier to train MAML++ than the original MAML. That paper also has a bit nicer code structure from the first glance. I would start from it.
download code for MAML++, download omniglot data set, run experiment to reproduce the results of MAML++ paper.
while it’s being trained, scan through MAML++ and (optionally) MAML. Understand the code
design an experiment for online learning, maybe one reproducing with OML/ANML. adapt the code accordingly. Run training
Here are some further options:
design a multistage architecture an implement it
or convert the network into depth prediction or camera parameters estimation in online mode
There is also a Reptile paper with code that can be useful. It proposes some simplification of gradient descent steps from MAML.
Video with a review of several methods of continual learning: Continual Learning in Neural Networks by Pulkit Agarwal