A-GEM implementation

[JonasFrey96]

Hello,

thanks for all your work. Could you explain why the A-GEM numbers are so bad ?
I ran examples/gem_agem_mnist.py and the performance on smnist are off by a big margin compared to GEM even when increasing the memory capacity.

Thanks a lot in advance.

Best regards from Zurich,
Jonas

[AndreaCossu]

Hi @JonasFrey96 😄 From our experiments, GEM works quite well with single head models, while A-GEM obtains poor performances. With multi-headed models, A-GEM recovers part of the performance in terms of catastrophic forgetting. This is the case also for Synaptic Intelligence: the original paper works with multi head models so the strategy performs poorly with a single head.
If you have been able to make A-GEM work with single-head , please let us know. That could indicate a bug in our implementation. However, up to now we have no sign pointing in that direction.

[JonasFrey96]

Thanks a lot for you quick response. I will investigate this a little bit further in the upcoming days. I will check the original Tensorflow implementation and let you know if i have more information.

[AndreaCossu]

Great, thanks!

[vlomonaco]

Moving this issue to "discussions", please @JonasFrey96 mark it as "answered" when you information need is satisfied! :)