About the meta learner

[Sherlock1118]

About the meta learner

[zyang1580]

Thank you!
To some degree, the work is inspired by MAML(optimized-based). Like MAML, it takes bilevel optimization. But, it also belongs to the Black-Box type (Model-based), because the transfer component of the framework can be thought of as a hyper-network (meta-model), which will generate recommender parameters for the next period. The above are my opinions.
Does the question "which work before give this spike to you" refer to which works inspire us? MAML and LambdaOpt.

[Sherlock1118]

thanks for your reply！！As far as I understand，the “transfer” component is a model-agnostic meta-leaner，which means it doesn't care about the specific forms of the base learner, and the meta learner doesn't need to be the same as base learner. In this way, can I say it is a more universal approach than MAML? And can the "transfer" be used in other areas(not only in recommender system)？

Thank you!
To some degree, the work is inspired by MAML(optimized-based). Like MAML, it takes bilevel optimization. But, it also belongs to the Black-Box type (Model-based), because the transfer component of the framework can be thought of as a hyper-network (meta-model), which will generate recommender parameters for the next period. The above are my opinions.
Does the question "which work before give this spike to you" refer to which works inspire us? MAML and LambdaOpt.

[zyang1580]

I agree with your understanding. Regarding the first question, I am not an expert in the area of meta-learning, so I am not sure whether it is more universal. I think MAML is limited if only the initial parameters can be meta-learned. Meanwhile, in some scenarios where MAML can be utilized, the proposed method can't be utilized. Regarding the second question, I think the answer is Yes. Sherlock ***@***.***> 于2021年4月2日周五 下午4:38写道：

…

Thanks for applying！！As far as I understand，the “transfer” component is a model-agnostic meta-leaner，which means it doesn't care about the specific forms of the base learner. In this way, can I say it is a more universal approach than MAML? and can the "transfer" be used in other areas(not only in recommender system)？ Thank you! To some degree, the work is inspired by MAML(optimized-based). Like MAML, it takes bilevel optimization. But, it also belongs to the Black-Box type (Model-based), because the transfer component of the framework can be thought of as a hyper-network (meta-model), which will generate recommender parameters for the next period. The above are my opinions. Does the question "which work before give this spike to you" refer to which works inspire us? MAML and LambdaOpt. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#3 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMO3KUZMY33L3KJYQSPKKE3TGV663ANCNFSM42GT6GBA> .

[Sherlock1118]

I agree with your understanding. Regarding the first question, I am not an expert in the area of meta-learning, so I am not sure whether it is more universal. I think MAML is limited if only the initial parameters can be meta-learned. Meanwhile, in some scenarios where MAML can be utilized, the proposed method can't be utilized. Regarding the second question, I think the answer is Yes. Sherlock @.***> 于2021年4月2日周五 下午4:38写道：
…
Thanks for applying！！As far as I understand，the “transfer” component is a model-agnostic meta-leaner，which means it doesn't care about the specific forms of the base learner. In this way, can I say it is a more universal approach than MAML? and can the "transfer" be used in other areas(not only in recommender system)？ Thank you! To some degree, the work is inspired by MAML(optimized-based). Like MAML, it takes bilevel optimization. But, it also belongs to the Black-Box type (Model-based), because the transfer component of the framework can be thought of as a hyper-network (meta-model), which will generate recommender parameters for the next period. The above are my opinions. Does the question "which work before give this spike to you" refer to which works inspire us? MAML and LambdaOpt. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#3 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMO3KUZMY33L3KJYQSPKKE3TGV663ANCNFSM42GT6GBA .

Thanks for your time！I have benefited a lot！