It has been claimed that for the performance it doesn’t really matter how the Fisher is computed. But while this holds to some extent for Split MNIST, already with Split CIFAR-10 significant differences in performance emerge.

At ICLR? Come and hear more at poster #483 on Saturday-morning!

Try out the class-incremental learning with repetition benchmarks of the challenge yourself! github.com/ContinualAI/...

These winning strategies clearly outperform experience replay on data streams *with* repetition, but on a “standard” task-based continual learning stream *without* repetition, experience replay performs better.

A striking outcome of the challenge was that all winning teams used some kind of ensemble-based approach, in which separate sub-networks per task/experience are learned and later combined for making predictions.

Thanks Simon! I’d be keen to be added as well ☺️