So the insightful part isn't that smart selection works but rather the use of clustering as a filtering mechanism. Using it removes the need for an expensive oracle - which is a huge Dub
Got it, thanks. I went somewhat deep into AL for my PhD so it kinda sounds obvious to me. Clustering is something you do as a side thing when applying e.g. entropy-based selection because otherwise you tend to pick outliers. But I guess if it where that obvious, someone would have done it before, so there's surely details I'm missing. Gotta check the paper 😁
The innovation is here is to take two synergestic ideas - self-supervised learning and active learning based on clustering- and combining them. Rather than present a whole new idea- this is more about presenting a very compelling solution together.
Also, I'm sure others have had the idea before. Meta just happened to publish it in this manner
Excellent summary! Question, through. Isn't this just the same argument the Active Learning community has been hammering all along?
So the insightful part isn't that smart selection works but rather the use of clustering as a filtering mechanism. Using it removes the need for an expensive oracle - which is a huge Dub
Got it, thanks. I went somewhat deep into AL for my PhD so it kinda sounds obvious to me. Clustering is something you do as a side thing when applying e.g. entropy-based selection because otherwise you tend to pick outliers. But I guess if it where that obvious, someone would have done it before, so there's surely details I'm missing. Gotta check the paper 😁
The innovation is here is to take two synergestic ideas - self-supervised learning and active learning based on clustering- and combining them. Rather than present a whole new idea- this is more about presenting a very compelling solution together.
Also, I'm sure others have had the idea before. Meta just happened to publish it in this manner