Few-Shot learning and Continual Learning

1 Like

You could be interested in our paper as well, in which we promote Continual-Meta Learning when you are interested in agents maximising their online rewards/performance they can cumulate when deployed.

Online Fast Adaptation and Knowledge Accumulation: a New Approach to Continual Learning

1 Like

Interesting recent paper on the matter: Defining Benchmarks for Continual Few-Shot Learning :smiley:

1 Like

Hi Vincenzo, I am one of the authors of the paper.

We are particularly interested in the intersection of few-shot and continual learning, and we have been following your discussions on slack and here.

It would be nice to have a deeper discussion (and a feedback) at some point. In particular, we are improving the paper with additional baselines. We have tested few-shot algorithms in this new setting but we would like to include some significant continual learning algorithm.

2 Likes

Hi Massimiliano,

I skimmed your paper real quick, will read it fully later. But it seems like you are trying to push CL in a similar direction as I am :slight_smile:

I think, however, that you are missing some related work on Continual Few-Shot learning.
Specifically, [1] introduced Continual Few-Shot learning. Then, [2] proposes a Continual Few-Shot learning algorithm that can detect the task boundaries. Finally, we really recently propose a new Continual Few-Shot learning setting, which we believe to be more aligned with real-life applications.

It would be nice if a section of your paper could explain how your new setting differs from these :slight_smile:

If you do, and you need help, lmk! I know the specifics of those settings really well.

[1] Task Agnostic Continual Learning via Meta Learning
Xu He, Jakub Sygnowski, Alexandre Galashov, Andrei A. Rusu, Yee Whye Teh, Razvan Pascanu

[2] Continuous Meta-Learning without Tasks
James Harrison, Apoorva Sharma, Chelsea Finn, Marco Pavone

[3] Online Fast Adaptation and Knowledge Accumulation: a New Approach to Continual Learning
Massimo Caccia, Pau Rodriguez, Oleksiy Ostapenko, Fabrice Normandin, Min Lin, Lucas Caccia, Issam Laradji, Irina Rish, Alexandre Lacoste, David Vazquez, Laurent Charlin

Thank you for letting us know of those papers. I am aware of the first one quite well. I’ve only skimmed the second one, just once, and was not aware of the third one. We’ll go over them carefully and update our paper. Thanks for offering to help. We will get back to you if we feel like a discussion would be warranted (we don’t want to waste your time otherwise). :slight_smile:

1 Like

please, don’t feel like you are wasting my time :slight_smile:

I’m happy to help! Continual Learning + Meta-learning FTW

Related post here: [ContinualAI Reading Group]: "Defining Benchmarks for Continual Few-Shot Learning" :smiley: