Logo

Balaji Lakshminarayanan ML talk - Shared screen with speaker view
vinay
33:40
Dear All,We are working on an OOD-Bible titled "Oodles of Odds".This is a work-in-progress and we plan to disseminate the paper this coming week.Here's the project URL:https://matthew-mcateer.github.io/oodles-of-oods/
vinay
37:29
Also, should you be using MNIST in your OOD-experiments, we have curated two 'semantically relevant' OOD datasets that you might find useful:- https://www.kaggle.com/c/Kannada-MNIST- https://github.com/Daniel-Wu/AfroMNIST
Mark Chen
46:28
Follow up question: Do you consider a blank image to be OOD? Almost every generative model will assign extremely high logp to a blank image.
Andrew Ross
46:52
I’m assuming all of the results are asymmetric, right? I.e. training on SVHN won’t lead to high log p(x) on CIFAR, and training on MNIST won’t lead to high log p(x) on Fashion MNIST?
Boaz Barak
48:39
@Andrew - good question - I don't want to stop Balaji now but please hold on to it and ask it
Boaz Barak
49:50
Answered now :)
Andrew Ross
50:19
Yep :)
Andrew Ross
01:11:03
How much easier does this problem become if you consider the problem of detecting whether a *batch* of samples is iid from the training distribution, rather than an individual example?
Andrew Ross
01:11:14
(Maybe a question for the end)
Andrew Ross
01:15:00
(I guess you’re getting to it right now :)
Boaz Barak
01:16:59
@Andrew you seem to be good at bits/dimension compression of this talk :)
Andrew Ross
01:17:05
Hehe :)
M. Wasil Wahi-Anwar
01:31:56
When you say ensembles consistently perform the best, are all the models in the ensemble trained on the exact same training data, or different data distributions?
John
01:33:46
Can you clarify what kind of ensembles are you referring to? Are they simple MLPs, ConvNets, other common architectures? Also, when training ensembles is that simple standard training, just SGD and probably some regularisation but nothing else?
Andrew Ross
01:39:11
Have you tried using Anchored Ensembles (https://arxiv.org/pdf/1810.05546.pdf) in such experiments?
Boaz Barak
01:39:56
I think he's close to the Q&A part so let's wait till then?
Puneet Dokania
02:04:08
Please continue if possible
Andrew Ross
02:04:10
Thanks for the great talk!
M. Wasil Wahi-Anwar
02:04:18
Thank you for having this!! This was great.