Andrea Cossu

COPE does not work anymore

1

The COPE strategy is not working on the Split MNIST benchmark on which it was originally tested (it achieves a performance around 20% accuracy instead of 94%). After some experiments,...

Add a plugin to manage the weight importance computation and penalization to the loss

1

By using a separate plugin to compute importances and penalize the loss we can reuse the same interface for EWC, MAS, SI and all future strategies.

Implement task descriptor Module

2

It would be nice to have a plugin that allows to concatenate to each input pattern a task descriptor. The task descriptor may be a simple one-hot task vector or...

Class order in W&B Confusion matrix

3

W&B confusion matrix should let the user specify a custom class order used when visualizing the confusion matrix. It is sufficient to set a boolean parameter similar to the one...

In our Zotero group, section `Benchmarks`, we should add references to benchmarks like Split MNIST / Omniglot and so on.

Reproduce AR1* results

3

[Link to the paper](https://openaccess.thecvf.com/content_CVPRW_2020/html/w15/Lomonaco_Rehearsal-Free_Continual_Learning_Over_Small_Non-I.I.D._Batches_CVPRW_2020_paper.html)

Reproduce CWR* results

1

[Link to the paper](https://openaccess.thecvf.com/content_CVPRW_2020/html/w15/Lomonaco_Rehearsal-Free_Continual_Learning_Over_Small_Non-I.I.D._Batches_CVPRW_2020_paper.html)

Experiments failed to be reproduced

1

Some tests failed. Check server log for details.

Some tests failed. Check server log for details.