Author Interview - VOS: Learning What You Don't Know by Virtual Outlier Synthesis

[@YannicKilcher]

Watch the paper review video here: kzbin.info/www/bejne/n16tZYdpqrF2b68 OUTLINE: 0:00 - Intro 2:20 - What was the motivation behind this paper? 5:30 - Why object detection? 11:05 - What's the connection to energy-based models? 12:15 - Is a Gaussian mixture model appropriate for high-dimensional data? 16:15 - What are the most important components of the method? 18:30 - What are the downstream effects of the regularizer? 22:00 - Are there severe trade-offs to outlier detection? 23:55 - Main experimental takeaways? 26:10 - Why do outlier detection in the last layer? 30:20 - What does it take to finish a research projects successfully? Paper: arxiv.org/abs/2202.01197 Code: github.com/deeplearning-wisc/vos

[@brandom255]

I think this new format with 1. Yannic explains paper (1½. Authors watch Yannic's video) 2. Common discussion is perfect! I really love it!

[@EternalKernel]

Love this format, this brings so much more value to us. Thank you!

[@deepblender]

In my view, quantifying uncertainty is going to play a more and more important role. The core idea of this paper to introduce uncertainty as a regularization loss is unexpected and very creative! Love it!

[@markmorgan5568]

Agreed.

[@markmorgan5568]

This was great. I plan to try it on a very practical problem I’m facing at work. I was glad to see that the “Gaussian” objection turned out to be not really an objection at all - they weren’t saying the feature distribution was Gaussian, just that the method implicitly makes that assumption. One thing I’d really like to hear (these researchers, especially!) thoughts on is whether we might be able to use autoencoders to identify OOD samples. The basic idea being that if a set of features can be reliably reconstructed after passing through a learned bottleneck, those features are likely to be “typical”, which in my mind is basically the same as “in distribution”. And if the reconstruction error is large, then the features are likely OOD. It seems like this would be true at any layer. I’d honestly be interested in thoughts from anyone reading this - I do think this is a reasonable approach, but I had limited success with it and am not sure if there’s a fundamental problem with the approach or if I probably just did something wrong.

[@phaZZi6461]

good insights, i was looking forward to this interview!

[@alan2here]

Image classifiers in self driving cars could probably do with broad backup categories: solid mover tumbleweed agent animal large animal small animal small spikey animal flying animal With score where applicable for "run", "freeze", "attack", "block", "run on lights", "run on horn", "freeze on lights", "freeze on horn", "car needs to flee", "car needs to close windows".

[@Addoagrucu]

last bit about starting the loss after the embedding space starts to take shape, reminds me of KL annealing in training a VAE

[@oncedidactic]

In retrospect the purpose and techniques of this paper are obvious in their practical utility and theoretical basis. But that always happens after many ideas have been tried and then the “a ha” approach pops out. This really gets me thinking about the other system 1 vs 2 and inter- vs extrapolation conversation going on at MLST for quite a while now. It seems like this outlier or uncertainty measurement will be of course necessary for combining NN (system 1 perception/“autopilot decisions”) with a system 2 process, which is evidently both harder and more expensive. Because you want to offload to system 1 as possible, and save system 2 for when it will pay off with knowledge/policy discovery. But to do that you need to know when system 1 is not giving a good result or is not applicable. Humans seem to have this capability (when they choose to exercise it, haha) and can fluidly extend and recombine system 1 and 2 as best fits the problem. A botanist can immediately tell the difference between “indistinguishable” weeds at 20 meters because they’ve trained themselves to have a plant detection system 1 eye-brain, by focusing on these data. The approach in this paper is a first step for making OD algo a moose-ologist in the same vein.

[@ChocolateMilkCultLeader]

Great interview. Lovely to see your work getting attention and recognition.

[@johnpope1473]

Well done. Great questions and answers.

[@zhenfang174]

A very good work!

[@shawnlee6633]

Kudos to Sharon!

[@HD-Grand-Scheme-Unfolds]

Hello Yannic Kilcher, always a pleasure and benefits watching your channel. Please though may I ask: With "VOS" does this means that it assists against the problem with insufficient robustness to adversarial attacks?

[@ofirshifman6472]

Great paper and interview, thanks! Did you ended up asking them about your comment on the p(x,b,y) notation? where in your first video you said you think there is a typo. I'm asking since I felt the same reading the paper and I'm trying to figure it out. I would like an answer if you know it :-) Thanks again and keep on the good work!

[@YannicKilcher]

Yes pretty sure it's a typo :)

[@MASTER-qc3ei]

Cat is all you need

[@josephharvey1762]

Just want to let you know, Yannic, in case you haven't noticed, the camera is focusing on your mic and not on you.

[@abdjahdoiahdoai]

#suggestion maybe ask what they are working on,

[@qichaoying4478]

Mr Cat distracted me for several times :(

[@ketalesto]

I love the format of two videos. Although, I think now both could be shortened a bit in length. Maybe shorten the paper review video and leave it at about 20 minutes, leaving some details for the authors to explain them. Just as a suggestion. :)

[@black-snow]

I object. He already left out a lot, didn't really go into details of the evaluation and didn't even look at the appendix. Not that it was bad, it was great. I just don't think the review could be any shorter without loosing essential information.