Can you please explain what you are trying to do with this attenution before the inner loop?

[AyanKumarBhunia]

        # Attenuate the initialization for L2F
        if self.args.attenuate:
            task_embeddings = self.get_task_embeddings(frames, task_id, names_weights_copy)
            names_weights_copy = self.attenuate_init(task_embeddings=task_embeddings,
                                                     names_weights_copy=names_weights_copy)

1. Can you please refer to some paper from where you took the inspiration regarding attenuation?

2. Can you please comment on how much the performance would drop if we do not use that and just follow off-the-shelf MAML?

I read a paper regarding this attenuation technique in 'Learning to Forget for Meta-Learning', CVPR 2020 (probably you all belong from the same research group).

Thanks?

[baiksung]

Hi,

Thanks for the interest in our paper.

Yes, as you have said, 'Learning to Forget for Meta-Learning' came from our lab, and I'm the first author.

1. You can find the motivation and methodology details on attenuation from the paper you mentioned. If you have more questions on this, you can directly e-mail me (dsybaik@snu.ac.kr) or you can open an issue on its github page:
   https://github.com/baiksung/L2F

2. The results reported on the paper are produced with the original MAML methodology. We have included L2F (attenuation) implementation as well just for the reference.

Hope this answers your questions.

[myungsub]

Hi,

as @baiksung said, the reported results are all from the original MAML, and we've added the additional meta-learning algorithms (L2F, MAML++, Meta-SGD) just for reference.

Thanks!