Ozan Sener (@ozansener) / X

Ozan Sener

3,948 posts

Ozan Sener

@ozansener

Loves Mathematics, Machine Learning, Juggling and People. Researcher at  and a Nerd (xkcd.com/356/). Previously: @StanfordAILab, @Cornell, and @IntelAI.

Munich 🥨

Joined December 2008

Ozan Sener
@ozansener
Nov 10, 2024
🚀 We are hiring full-time ML researchers and PhD-level interns! Join us for exciting projects in AI for Science (weather and climate models) and understanding LLMs. Full-time role: jobs.apple.com/en-us/details/… lists US but EU is also OK—apply regardless of location preference!
83K
Ozan Sener
@ozansener
Apr 26, 2020
Recently, I got interested in a new problem, derivative-free optimization. It is also called random search. We are presenting two papers on DFO in the next two days in #ICLR2020. (Note: this work was largely inspired by @beenwrekt's beautiful work: arxiv.org/abs/1803.07055.)
Ozan Sener
@ozansener
Feb 12, 2024
Replying to @francoisfleuret
Not really a measure theory person but I think the fundamental problem is axiom of choice allowing construction of some pathological subsets. Example:
en.wikipedia.org
Vitali set - Wikipedia
4.4K
Ozan Sener
@ozansener
May 21, 2018
Replying to @BarbaraCoastal and @dhh
In CA, if you make around 100K, you are taxed like 30 percent. Americans are still paying lots of tax but not getting anything back
Ozan Sener
@ozansener
May 12, 2020
Replying to @MaartenvSmeden
Machine learning orders a beer, statistics pay for it and AI drinks it.
Ozan Sener
@ozansener
Apr 30, 2017
Awesome paper by @stanfordnlp arxiv.org/pdf/1704.06956… We definitely need sth similar for robot learning and grounding robotic instructions
Ozan Sener
@ozansener
Jun 17, 2020
This is a work I really like and proud of so here is a little bit more context for it. Most generalization bounds depend on the dimension of the parameter space. This is very problematic especially for neural networks as dimension is typically large (in millions)...
Umut U. Simsekli
@umutsimsekli
Jun 17, 2020
Replying to @umutsimsekli
Our theory essentially tells: "SGD generates fractals and the gen. err. is bounded by the intrinsic dimension of these fractals, which is determined by the tail-behavior at convergence". The paper: bit.ly/2Y91ddR Joint w/ @ozansener @MuratAErdogdu @GeorgeDeligian9 2/2
Ozan Sener
@ozansener
Jul 26, 2018
One of my #NIPS2018 reviews is 247 characters long, they could have just tweeted instead!!! And, their confidence score is 5 just as I expected.
Ozan Sener
@ozansener
Nov 27, 2017
My chat with the uber driver who picked me up from campus -UD: what do you work on +ME: Machine learning -UD: What do you think about Hinton's CapsuleNet? #OnlyInBayArea
Ozan Sener
@ozansener
Jun 14, 2018
Replying to @fhuszar
Doesn't work in bay area though. Once an Uber driver asked me sth similar and I answered as I am doing statistical learning. He then asked what do you think about Hinton's Capsule Networks.
Ozan Sener
@ozansener
Nov 10, 2019
Replying to @sibirbil
Yuksek lisansin bu durumu kokunden cozmek icin cok gec kalmis bir yer oldugunu dusunuyorum. Akademi/toplum/aile insanlara basarilari kadar saygi duydugu ve deger verdigi surece, insanlar da kendilerine basarilari kadar deger verecek ve saygi duyacaklardir.
Ozan Sener
@ozansener
Sep 4, 2018
Two papers have been accepted to #NIPS2018. One is on using multi objective optimization for multi-task learning (pre-print coming soon); and, other one is on domain generalization via distributional robustness (arxiv.org/abs/1805.12018). See you all in Montreal.
arxiv.org
Generalizing to Unseen Domains via Adversarial Data Augmentation
We are concerned with learning models that generalize well to different \emph{unseen} domains. We consider a worst-case formulation over data distributions that are near the source domain in the...
Ozan Sener
@ozansener
Mar 6, 2019
I heard the sentence "...Every matrix corresponds to a graph..." first time during the first day of my PhD from @DavidBindel in Matrix Computations class. It was mind blowing. I am sure this blog post will be similarly mind blowing to many people.
Tai-Danae Bradley
@math3ma
Mar 6, 2019
A nice fact I like: Every matrix corresponds to a graph, and so familiar things (e.g. matrix multiplication) have nice pictures! Another nice fact: joint probability distributions *also* correspond to graphs. They have telling pictures, too. New blog post! math3ma.com/blog/matrices-…
Ozan Sener
@ozansener
May 1, 2018
If you want to learn how to do active learning for CNNs, come checkout our poster (#38) at #iclr2018 today between 4.30pm and 6.30pm. Joint work with @silviocinguetta at @StanfordCVGL.