Who is the bald guy on the far right in the checkered blue shirt? He asked an important question: what is the agreement (for the original ImageNet pictures) between the OLD ImageNet labels created through the 2011 process and the NEW labels created through the 2019 process?? This is not in the paper! Also where are the slides? They don't seem to be on Ben Recht's website or the Simons website.
I see what he’s saying, but I’m not fully convinced that it’s distribution shift that’s causing the problem. I mean, to be clear, it is “distribution shift” in the sense that the data is different, but I don’t think the distributions have shifted in ways that *should* be measurable. I think it’s evidence that the things we are getting the classifiers to learn are not causally connected to reality. That’s an architecture problem, not a distribution shift problem. Also, I feel like this guy seems nice at first, but is being quite rude. I hope other people aren't encouraged to behave like him in the future, we shouldn't treat the people asking questions of our work this way!! That's counterproductive for science.
Who is the bald guy on the far right in the checkered blue shirt? He asked an important question: what is the agreement (for the original ImageNet pictures) between the OLD ImageNet labels created through the 2011 process and the NEW labels created through the 2019 process?? This is not in the paper!
Also where are the slides? They don't seem to be on Ben Recht's website or the Simons website.
I think its Nathan Srebro.
@@samlaf92 thank you!
I see what he’s saying, but I’m not fully convinced that it’s distribution shift that’s causing the problem. I mean, to be clear, it is “distribution shift” in the sense that the data is different, but I don’t think the distributions have shifted in ways that *should* be measurable.
I think it’s evidence that the things we are getting the classifiers to learn are not causally connected to reality. That’s an architecture problem, not a distribution shift problem.
Also, I feel like this guy seems nice at first, but is being quite rude. I hope other people aren't encouraged to behave like him in the future, we shouldn't treat the people asking questions of our work this way!! That's counterproductive for science.
well said!
For the most part ML isn't really science, is the vibe I get from him, which I agree with. He doesn't seem rude at all to me