NeurIPS 2023 Poster Session 1 (Tuesday Evening)
Вставка
- Опубліковано 12 гру 2023
- Papers:
CamoPatch: An Evolutionary Strategy for Generating Camoflauged Adversarial Patches (openreview.net/forum?id=B94G0...)
FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning (arxiv.org/abs/2309.14062)
Causes and Effects of Unanticipated Numerical Deviations in Neural Network Inference Frameworks (openreview.net/forum?id=6zyFg...)
Conservative State Value Estimation for Offline Reinforcement Learning (arxiv.org/abs/2302.06884)
Links:
Homepage: ykilcher.com
Merch: ykilcher.com/merch
UA-cam: / yannickilcher
Twitter: / ykilcher
Discord: ykilcher.com/discord
LinkedIn: / ykilcher
If you want to support me, the best thing to do is to share out the content :)
If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: www.subscribestar.com/yannick...
Patreon: / yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n - Наука та технологія
Hit them with the classic “What’s your justification for assuming Gaussian data?” 😂
Kala mo naman may alam ka
Ruthless
I mean, you always have to assume something, I liked the way he was able to explain the rationale. Seems like a pretty robust work in comparison to others. I just wonder how no one implemented it before. I have to be honest, I am rather sick of seeing work that does nothing but explain some complicated math (hack the reviewer, there is one country in which all of the students do it, guess which), at least this one made sense.
We did a random projection, by the central limit theorem the distribution is now normalish… probably me IRL.
@@sundial3914 "accidentally" forgetting the IID part
Yannic, cool. Please more contents about this conference.
Great stuff Yannic, love this type of content 😊. Posters session is the thing I love the most about conferences!
Glad to see this in it's entirety.
My colleague's told me about your channel, and your content and commentary is amazing
What a great format! Lot's of ♥ to everyone involved.
Exemplar free learning sounds cool. I'm interested to see how we can train in the absence of persistent training data, or when training data is ephemeral.
Thank you for sharing.
Kudos for keeping the production quality quite high while being at the conference!
Thanks, feels like being there :)
Very appreciated this video!
Stuff like this makes me regret dropping out of uni a little. So much cool stuff to learn.
I've dropped out but I keep doing research and writing papers on my free time. I'll be there next year. 🎉
Keep doing what you love!
Yeah. I went to ICCV in 2013 for my work, and I remember the poster section being my favorite part of it. So much cool stuff.
I'm too, and regrets that my bachelor and master degree works were not so high quality - without conferences and papers or other science things
uni = learn ???? wtf
Love this!
great ideas on poster number 2 on continual learning. It is a huge open problem.
Appreciate the citation for the "useless" of statistical determinism. Growing concern on mine is that we may be pushing our "good enough," computing to its limits. With safety being something that cannot be guaranteed.
Another area where this impacts my own work is branch prediction in a graph programming model. Previous job I had ran into that issue within the scope of authentication being skipped over in a complex computation environment.
My concern here is the combination of floating point Rounding preventing deterministic output. With the addition moving some of those calculations into logically determined graph program. Feels like we are being sandwiched into a strange situation.
My interest would be if we were to focus on risc-v and gain access to the mythic slow good computation model. If we could scale that system. That would be the safest bet. But don't see any momentum leading there.
dont worry about first 3.
and youre right about #4, its not going
This is a great video
Cool! It's like I'm there! Apart from you are actually talking to people :)
5:24 was good thanks
It was a good paper!
So fun!!
The flirt is strong with this one.
The last poster about offline RL is cool.
Would love to see MLST have another NeurIPS season
Love Mahalanobis distance , very under utilized
Uni innsbruck: interesting findings. I did not know it makes differences at all on CPU and GPU hardwares, with huge error potential, surprise, surprise. If we consider that we are using 64bit system, the error on 7th decimal point is visible, then it will be significant errors - iteration in DNN.
I think the Masked Adversarial Patches are a cool idea but the fact that they greedily target the most vulnerable piece of an image means it's far more difficult to employ them in the real world...
Originally the whole point of them was that you could place them *anywhere* within image - and then even going so far as to working when distorted (i.e. when seen at fairly arbitrary angles in a possibly free-moving camera from fairly far away and fairly close up and rotated etc.) and so I'd argue this kind of adversarial patch is indeed "unrealistic", at least with those applications in mind.
As is, by the given explanation, this isn't necessarily that different from the works that consider the whole image but only allow modifying a sparse set of pixels, say.
Thanks for your insightful explanation!
🎉🎉
@yannic Can you hunt down the Mamba people?
So, were those all examples of what you described as poor presentation at the beginning?
Do MSAtransformer paper
Where is it?
this also looks like a great place for Triumph the insult dog to visit 😏
That's a great diffusion model...for me to poop on!
bro will I see you there?
When going undercover involves NOT wearing sunglasses.
It’s funny that you introduce yourself, I mean who doesn’t know you ?! 😅
100% unrecognizable without sunglasses!
@@shaxosYT Right, his sunglasses are adversarial patches!!
I find something very human in this. What a lovely concept!
Where is George?
I can't believe Yannic has real eyes 😮 i thought he was born with cool sunglasses
Lirst💜👌💯✨️🔥👽
no glasses🕶️？！
Aren't we moving past paper, especially in the extreme dimensionality and temporal domains of AI. Audio and video doesn't work so well on paper.
How far are we from a shot for shot remake of a movie? 18 months?
about two more papers
bro where are the sunnies , its like I dont even know who you are ?
The reverse Clark Kent effect is real. Who is this weird guy who sounds like Yannic?
8:00 - Pretty obvious what happened here, my man throws down some serious game while filming the original session. Fast forward to the next morning and, like an absolute champ, la chica here records a fresh audio track that doesn’t include the bits Yannic was spitting fire. I will say, you nearly gave yourselves away when you started tickling her at the end like that but Nora pushed through. Bravo you two 👏 😅
Zac the joker 😂😂
First
Where is George?