Khue Le @netw0rkf10w Twitter profile

Last Seen Profiles

@morexts89

@triss808

@Jalpha98

@NorthernArgus

@bokeplokalmalam

@FHVallejo

@achilleansappho

@alshaaeirana

@RealClpz

@stw_pdg

@marcoghisla

@Milan5i

@G7

@Kimsoohyunz

@bedniy_evrey

@bokeplokalmalam

@icouldeatyou_

@x2bucky

@BeyondThePosts_

@gyungsik_

@MatthewRockmore

@jmertic

@NU_FAME

@PasutriAsli2

@FR9QG

@Jane_marumaru

@indy_traffic

@eigengraux_

@lisandroganuza

@info_kls

@bokeplokalmalam

@BinorRaja

@kurtmac

@keek_art

@theothertobii

@Mbasitpml

Khue Le

@netw0rkf10w

5 months

Hi @aaron_defazio . Here's the result of my optimizer, compared to yours (still running). Can you beat my blue curve with hyper-parameter tuning? ;) Please give it a try using this code:

Aaron Defazio

@aaron_defazio

5 months

Schedule-Free Learning We have now open sourced the algorithm behind my series of mysterious plots. Each plot was either Schedule-free SGD or Adam, no other tricks!

41

238

1K

4

10

125

Khue Le

@netw0rkf10w

5 months

While waiting for @aaron_defazio 's tuning result, here's my full run of his method (green curve). Interestingly, some modifications inspired by my optimizer seem to boost its performance. Note: MAE's default hyper-params are used for all experiments.

Aaron Defazio

@aaron_defazio

5 months

@netw0rkf10w Ok, I will try and run early next week.

0

21

8

7

61

Khue Le

@netw0rkf10w

5 years

This reminds me of this Flaubert paper: that gave proper credits to ULMFiT (by @jeremyphoward and @seb_ruder ).

2

18

43

Khue Le

@netw0rkf10w

2 years

@mblondel_ml This is actually not new. We (with @inthebrownbag ) published this result at NeurIPS 2021 in our paper on Frank-Wolfe for MAP inference (), in which we provide convergence results for the general non-convex and non-smooth settings. 1/2

1

5

31

Khue Le

@netw0rkf10w

5 months

@MrCatid @aaron_defazio Yes. The question is when. It really depends on the reaction of the community. If a lot of people are eager to try (as with Aaron’s method) then I’ll try to release it as soon as possible. Otherwise I’m afraid it’ll take time.

2

1

10

Khue Le

@netw0rkf10w

5 months

Summary of the rookie mistakes made by the Keras team:

0

2

8

Khue Le

@netw0rkf10w

4 years

@A_K_Nain Call it kerax!

0

9

Khue Le

@netw0rkf10w

5 months

@MrCatid @aaron_defazio Why should I release it early without a proper accompanying paper if nobody cares?

1

0

9

Khue Le

@netw0rkf10w

2 years

@sp_monte_carlo 'CCCP is Frank-Wolfe in disguise' is actually a known result. Please see here:

Khue Le

@netw0rkf10w

2 years

@mblondel_ml This is actually not new. We (with @inthebrownbag ) published this result at NeurIPS 2021 in our paper on Frank-Wolfe for MAP inference (), in which we provide convergence results for the general non-convex and non-smooth settings. 1/2

1

5

31

0

3

8

Khue Le

@netw0rkf10w

7 years

Our paper "Tight Continuous Relaxation of MAP Inference: A Nonconvex Perspective" has been accepted for publication at CVPR 2018! With Prof. Nikos Paragios @AgoniGrammi . PDF available soon. Code available prior to the conference. #CVPR #CVPR2018

0

2

8

Khue Le

@netw0rkf10w

7 years

CVPR 2018 received 3303 valid submissions (all-time record of #CVPR ). Acceptance rate: 979/3303 ~= 29.6% #CVPR2018 Last year's acceptance rate: 782/2680 ~= 29.2% #CVPR2017

0

2

7

Khue Le

@netw0rkf10w

5 years

Hmm... I cannot agree with @jeremyphoward on some points here: 1. LSTM was not created by Schmidhuber, but by "Hochreiter and Schmidhuber". and "Perhaps most important of all": 2. Back-propagation was invented by Seppo Linnainmaa. See

1

0

5

Khue Le

@netw0rkf10w

5 years

@ptrblck_de @PyTorch @soumithchintala This is amazing! The PyTorch community is lucky to have you. Thanks for your dedication!

0

3

Khue Le

@netw0rkf10w

2 years

@denny_zhou To be able to say that “infinity is not the largest number”, one is required to define “number”, “infinity”, and “larger”. When working this out one will see that it’s not that simple. Both GPT3 and the kid can be right or wrong, depending on the definitions.

0

5

Khue Le

@netw0rkf10w

3 years

@gentaiscool @seb_ruder @emnlpmeeting Congratulations! I would like to point out very relevant work on a French benchmark called FLUE that appeared on arXiv on 11 December 2019: . (An evaluation server is also under construction: .)

0

1

4

Khue Le

@netw0rkf10w

4 years

@gpapamak @mpd37 @kchonyc @ChengSoonOng It seems you are describing the *derivative* instead of the *gradient*. I agree they are conceptually different. (After all it's just a convention so I have no problem accepting both, but I wouldn't say one is more correct than the other.)

1

0

4

Khue Le

@netw0rkf10w

5 months

@aaron_defazio Looks very promising, congratulations! Would you mind sharing some results for ViT on ImageNet?

1

0

4

Khue Le

@netw0rkf10w

5 months

@omead_p Yes there will be a paper, in a month or so, I hope. The code may be released sooner.

1

0

4

Khue Le

@netw0rkf10w

2 years

@mblondel_ml @inthebrownbag Here you go!

0

4

Khue Le

@netw0rkf10w

2 years

In our NeurIPS 2021 paper (with @inthebrownbag ) we showed that CCCP is Frank-Wolfe in disguise. Happy to see other people recently rediscovering this fact and presenting it as a striking result. Want to know another equally striking fact? Mean Field is also Frank-Wolfe!👇

Khue Le

@netw0rkf10w

2 years

@mblondel_ml This is actually not new. We (with @inthebrownbag ) published this result at NeurIPS 2021 in our paper on Frank-Wolfe for MAP inference (), in which we provide convergence results for the general non-convex and non-smooth settings. 1/2

1

5

31

0

1

4

Khue Le

@netw0rkf10w

2 years

@mblondel_ml @inthebrownbag As a byproduct we also provide a justification for the unit step-size used in CCCP (c.f. screenshot above). Basically I can say that only Section 4 of the above paper is new, compared to our results. It's worth mentioning that Mean Field inference is also Frank-Wolfe! 2/2

1

0

4

Khue Le

@netw0rkf10w

3 years

@inthebrownbag and I are happy to present our recent work on regularized Frank-Wolfe for MAP inference at #NeurIPS2021 . You are welcome at our poster spot A0 () between 9:30am-11am CET tomorrow! Paper: Code:

1

4

Khue Le

@netw0rkf10w

3 years

@ccanonne_ IMO any non-standard result should be accompanied by a proof, at least as a footnote or an appendix. The above inequality can be proven in just 1 line (CS followed by sum >= summands), so a footnote would be appropriate in this case.

0

4

Khue Le

@netw0rkf10w

5 months

@aaron_defazio That's awesome, thanks! Looking forward to it.

0

2

Khue Le

@netw0rkf10w

1 year

Interesting ideas of using Optimal Transport for learning to align two sequences of features.

Hang Le

@formiel

1 year

Happy to share our recent work on speech translation (ST): , which will be presented at ICML 2023. We make two contributions: 1) showing that CTC can reduce the modality gap in ST pre-training... 1/2

3

5

29

0

3

Khue Le

@netw0rkf10w

2 years

@binhtngn @gabrielpeyre The connection to FW and the O(1/k) rate of convergence of CCCP had been shown previously in . See also this thread:

Regularized Frank-Wolfe for Dense CRFs: Generalizing Mean Field and Beyond

We introduce regularized Frank-Wolfe, a general and effective algorithm for inference and learning of dense conditional random fields (CRFs). The algorithm optimizes a nonconvex continuous...

arxiv.org

Khue Le

@netw0rkf10w

2 years

@mblondel_ml This is actually not new. We (with @inthebrownbag ) published this result at NeurIPS 2021 in our paper on Frank-Wolfe for MAP inference (), in which we provide convergence results for the general non-convex and non-smooth settings. 1/2

1

5

31

0

1

2

Khue Le

@netw0rkf10w

6 years

My work is featured in @RSIPvision , but not like the other ones. Not sure whether I should be happy or sad 😂😂😂

RSIP Vision

@RSIPvision

6 years

Visa refused @centralesupelec @inria @AgoniGrammi @RSIPvision @CVPR #CVPR #CVPR18 #CVPR2018 #ComputerVision #ArtificialIntelligence #AI #CVPRdaily

2

14

33

0

1

3

Khue Le

@netw0rkf10w

7 years

Neural Machine Translation (seq2seq) Tutorial

0

3

Khue Le

@netw0rkf10w

5 months

@aaron_defazio It's worth pointing out that PyTorch's default pre-trained ResNet50 achieved 80.858 top-1 accuracy (), so it seems your baseline is a bit weak. It could be interesting to see the results on a stronger baseline such as PyTorch's reference implementation.

1

0

3

Khue Le

@netw0rkf10w

6 years

What a bright future for AI in France: - Google is opening a new AI research lab in Paris. - DeepMind is opening a new AI research lab in Paris. - Facebook AI Research Paris is doubling their size. - Samsung is opening a new AI research lab in Paris. I am really excited!

0

1

3

Khue Le

@netw0rkf10w

8 years

⚡️ “Tensorflow and deep learning - without a PhD” by @martin_gorner

0

3

Khue Le

@netw0rkf10w

6 years

New MSc in Artificial Intelligence at @centralesupelec with excellent curriculum:

Paragios, Nikos

@AgoniGrammi

6 years

at the edge of academic excellence @centralesupelec introduces new #graduate #MSc on #ArtificialIntelligence Boost your career with introductory, excellence, domain-specific self-contained #AI curriculum Symbolic, subsymbolic, decisional #DataDriven #AI

0

12

22

0

3

Khue Le

@netw0rkf10w

5 months

@aaron_defazio Awesome. Thanks for sharing, Aaron. The good news for me is that it's not the same algorithm that I have been developing for a while (there's some similarity though). Will compare and share the results soon.

1

0

3

Khue Le

@netw0rkf10w

4 years

Want to learn about scalable machine learning? The #RaySummit @raydistributed has a great program with a great speaker lineup. It's about to start, but still not too late to register:

0

3

Khue Le

@netw0rkf10w

4 years

@gpapamak @mpd37 @kchonyc @ChengSoonOng Screenshot taken from . (This is not to show that Wikipedia is correct, just that there exists the mentioned convention.)

1

0

2

Khue Le

@netw0rkf10w

4 years

I always have a lot of respect for authors who make their books freely available. My hat’s off.

Sheldon Axler

@AxlerLinear

4 years

A new electronic version of Measure, Integration & Real Analysis is now available at . This new version dated 30 July 2020 fixes one more minor typo. The website above also has a list of all known typos in the printed version of the book. #measuretheory

0

2

27

0

2

Khue Le

@netw0rkf10w

2 years

An ICLR 2023 submission has been accused of being a rehash of previous work, claim supported by detailed technical arguments. If true then there must be consequences. Intentional misleading contributions should not be tolerated in academic research.

Git Re-Basin: Merging Models modulo Permutation Symmetries

The success of deep learning is due in large part to our ability to solve certain massive non-convex optimization problems with relative ease. Though non-convex optimization is NP-hard, simple...

openreview.net

0

2

Khue Le

@netw0rkf10w

4 years

@ducha_aiki @amy_tabb @tdietterich "Why not before?" You may think that double-blind reviewing is not necessary, but many people would disagree with that, including myself. Imagine publishing and marketing your submission during the rebuttal period as FAIR did with their DETR paper? Not a good thing to do. 2/2

1

0

2

Khue Le

@netw0rkf10w

5 months

@giffmana "multi-crop eval": Are you referring to "test-time augmentation"?

1

0

2

Khue Le

@netw0rkf10w

3 years

We would like to thank @Genci_fr for their support!

0

2

Khue Le

@netw0rkf10w

5 months

@danyallah_ @aaron_defazio Yes. I'll share the results gradually. Let me know if there's a specific problem that you want me to try on.

0

2

Khue Le

@netw0rkf10w

2 years

@mblondel_ml @inthebrownbag "vanilla FW on an epigraph reformulation" is **precisely** generalized FW. They are exactly the same. If you write Prob. (3.1) under form (3.2) with variable t, then iteration (3.3) can also be write back to the form without t, and that's generalized FW. Please let me know 1/2

1

0

2

Khue Le

@netw0rkf10w

4 years

@ducha_aiki @amy_tabb @tdietterich Obviously until the authors want to reveal their names, depending on their next submission. If a paper is on arXiv, people know it exists and can cite it, it cannot be scooped. Thus, I'm not sure your arguments about rejects are valid. 1/2

1

0

2

Khue Le

@netw0rkf10w

7 years

#CVPR 2018 reviews are out! Several things have changed this year: 4 reviews, different preliminary rating levels. What else?

0

2

Khue Le

@netw0rkf10w

4 years

Several European supercomputers working on COVID-19 research have been hacked to mine cryptocurrency. Please stay safe, @PRACE_RI @Genci_fr ! (cc @s_requena )

European supercomputers hacked to mine cryptocurrency

Several high-performance computers working on COVID-19 research have been forced offline following a string of attacks

www.welivesecurity.com

0

2

Khue Le

@netw0rkf10w

4 years

TIL @github has a great new feature: Discussions. Now any repo is basically an online forum. This is really awesome! Separating Discussions from Issues makes it much more easier to manage the content, especially for popular repos.

0

2

Khue Le

@netw0rkf10w

5 years

@s_requena @InriaParisNLP @CNRS This FlauBERT used your Jean Zay cluster :D

2

0

2

Khue Le

@netw0rkf10w

7 years

CVPR 2018 accepted papers: #cvpr #cvpr2018

0

2

Khue Le

@netw0rkf10w

5 years

After having been phenomenal in NLP, now it's time for self-supervised learning to conquer Computer Vision.

Oriol Vinyals

@OriolVinyalsML

5 years

Rapid unsupervised learning progress thanks to contrastive losses, approaching supervised learning! -40% Multitask SSL (2017) -50% CPC (2018) -70% AMDIM/MOCO/CPCv2/etc (2019) -76.5% SimCLR (2020, so far)

6

206

639

0

2

Khue Le

@netw0rkf10w

4 years

@fbk_mt @laurent_besacie Vietnamese 😍😍😍

0

2

Khue Le

@netw0rkf10w

7 years

A good blog post for understanding Faster R-CNN in detail. Also good for understanding key components of modern Object Detection systems. Faster R-CNN: Down the rabbit hole of modern object detection - Tryolabs Blog

0

2

Khue Le

@netw0rkf10w

4 years

TIL @IJCAIconf 's first conference was held in 1969! Probably the conference with the longest history of the field (all AI, vision, NLP included). The entire proceedings FROM 1969 are available for download: . IMPRESSIVE!

1

2

Khue Le

@netw0rkf10w

6 years

Last admission session of @centralesupelec 's MSc in Artificial Intelligence is on June 25th:

CentraleSupélec

@centralesupelec

6 years

Deadline to our last admission session of our MSc in #ArtificialIntelligence is approaching (June 25th)! ◀️◀️▶️▶️ #machinelearning #MachineIntelligence #deeplearning , #reinforcementlearning #decisionalAI #graphicalmodels #informationtheory

0

7

10

0

2

Khue Le

@netw0rkf10w

5 months

@Jacoed @aaron_defazio That’s already more than enough for this experiment. You can try it yourself if you want longer training, the code is here:

mae/FINETUNE.md at main · facebookresearch/mae

PyTorch implementation of MAE https//arxiv.org/abs/2111.06377 - facebookresearch/mae

github.com

0

2

Khue Le

@netw0rkf10w

5 months

@tienhaophung @aaron_defazio Started looking at this today but it's getting crazy. Training time is just too high (~40 days on 8 A100 GPUs??). And too many red flags in the code: no AMP, even no training resume (wtf?). Please suggest a more reasonable task with a better quality codebase.

1

0

2

Khue Le

@netw0rkf10w

5 years

@s_requena @InriaParisNLP @CNRS It is stated in the paper that the model was trained on (only) 32 GPUs for 410 hours. The authors will probably be able to provide more details on this. @formiel @laurent_besacie @didier_schwab

1

0

2

Khue Le

@netw0rkf10w

4 years

A nice summary of the (many) models available in the transformers library.

Sylvain Gugger

@GuggerSylvain

4 years

There are a lot of models in the transformers🤗 repo. Feeling lost? I know I was, so I made a little high-level summary of the differences between each model.

24

273

1K

0

2

Khue Le

@netw0rkf10w

7 years

@hbou @ylecun Les modèles sont publiés aussi:

1

2

Khue Le

@netw0rkf10w

4 years

@dennybritz Few-shot learning is hard. Good luck!

0

2

Khue Le

@netw0rkf10w

7 years

Semantic Segmentation using Fully Convolutional Networks over the years + PyTorch code

0

2

Khue Le

@netw0rkf10w

5 years

CVPR’20 decisions are available.

Greg Mori

@greg_mori

5 years

CVPR 2020 paper decisions are now available, 1470 papers accepted: A huge thank you to everyone involved -- 198 Area Chairs, 3664 reviewers, and the authors of the 6656 papers. #cvpr #ai #ML

7

93

426

0

1

Khue Le

@netw0rkf10w

6 years

@fhuszar @roydanroy @jeremyphoward @tdietterich @fchollet "he blocked the last time I disagreed with him" What?? 😮

1

0

1

Khue Le

@netw0rkf10w

7 years

Nice blog post about mixup, a recent data augmentation technique for neural networks.

Ferenc Huszár

@fhuszar

7 years

By popular demand: my thoughts on the mixup data-augmentation technique

13

73

274

0

1

Khue Le

@netw0rkf10w

7 years

@hardmaru It is currently faster than MathJax, and that’s all. @MathJax supports a larger set of commands/environments. And the upcoming MathJax 3 will have a significant boost in performance.

0

1

Khue Le

@netw0rkf10w

2 years

@mblondel_ml @inthebrownbag actually NOT that vanilla, sorry for the typo. 2/2

1

0

1

Khue Le

@netw0rkf10w

4 years

@ducha_aiki @amy_tabb @tdietterich Maybe I misunderstood it but you proposed "hands off arXiv", which is a "solution" (that is, for me, better than completely banning arXiv, so I think we agree on this). I'm glad that you also agree with my proposal ;)

1

0

1

Khue Le

@netw0rkf10w

1 year

@konstmish I forgot to include the reference: Apologies if I’ve overlooked something. @prof_grimmer

1

0

1

Khue Le

@netw0rkf10w

4 years

For comparison, CVPR started in 1983 but only digitalized proceedings of 1988 and later are available:

0

1

Khue Le

@netw0rkf10w

4 years

@ParcolletT @UnivAvignon I was essentially referring to the US system where there's a clear distinction between assistant/associate. I think the conversion is fair enough (in the sense that the typical order should be: PhD -> Assistant -> Associate --> Full). That's only a personal perspective though ;)

0

Khue Le

@netw0rkf10w

1 year

@LabrakYanis @ParcolletT Are you aware of FlauBERT? It’s concurrent work to CamemBERT and is highly relevant to your work.

1

0

1

Khue Le

@netw0rkf10w

6 years

2017 impact factor of computer vision journals! IJCV has surpassed PAMI!

Paragios, Nikos

@AgoniGrammi

6 years

(new) #ImpactFactor for #computervision , #patternrecognition & #imageprocessing journals (thanks to @ELS_Multimedia )

0

2

1

0

1

Khue Le

@netw0rkf10w

2 years

@theshawwn Could you share what your inspiration was? I have been working on this for more than a year (cc @inthebrownbag ), maybe we had the same inspiration ;) Actually what you described will not work well, it lacks an important step.

1

0

1

Khue Le

@netw0rkf10w

7 years

Elsevier in a nutshell: "I review for them for free, then I pay them to publish my paper, then my university pays them so I can read my own paper." (quote from some professor)

0

1

Khue Le

@netw0rkf10w

5 years

Warning: if you use @TensorFlow 2.0 and observe low accuracies, that's maybe because its ImageNet weights are broken:

resnet50 imagenet weights are differents in tf.keras vs keras · Issue #33459 · tensorflow/tensorflow

System information Have I written custom code (as opposed to using a stock example script provided in TensorFlow): no OS Platform and Distribution (e.g., Linux Ubuntu 16.04): win10, pycharm (same i...

github.com

0

1

Khue Le

@netw0rkf10w

4 years

@ducha_aiki @amy_tabb @tdietterich There's no contradiction. NLP conferences have an anonymity period of 1 month for a good reason. While strict double-blind is important, forbidding completely authors to publish their work on arXiv before submission have downsides (some of your post's arguments apply here) 1/2

1

0

1

Khue Le

@netw0rkf10w

4 years

Linear Algebra Done Right has been made freely available by Springer. This is a beautiful book in every sense.

0

1

Khue Le

@netw0rkf10w

5 years

@nikiparmar09 Could you please update the arXiv paper with the camera-ready version? The hyperlinks in the NeurIPS version are not clickable. Thanks.

0

1

Khue Le

@netw0rkf10w

6 years

@AgoniGrammi @essec @centralesupelec @MIT @UCBerkeley @Stanford @gchevil Excellent program! Last admission session deadline was on April 12th ;) Fingers crossed for my wife...

0

1

Khue Le

@netw0rkf10w

6 years

YOLOv3 is out (fast object detection system). The technical report is very informal, but easy and fun to read: Code:

0

1

Khue Le

@netw0rkf10w

6 years

@zacharylipton I saw a lot of errors like this on Google Scholar. I reported to them several times months ago but apparently those errors are still there. On their support page it is said that they cannot correct the errors “manually” :)

0

1

Khue Le

@netw0rkf10w

5 months

@jeankaddour @SunnySanyal9 @aaron_defazio One can study any existing techniques, but it should be made clear that those are existing. Averaging the last checkpoints, in particular, has been used extensively by the community (especially in NLP & speech, even before Transformer). Giving it a new name seems misleading IMO.

2

0

Khue Le

@netw0rkf10w

6 years

Semantic Image Segmentation with Tensorflow: Google DeepLab-v3+ official code and models

1

0

1

Khue Le

@netw0rkf10w

4 years

@fvsmassa @alcinos26 @szagoruyko5 Nice codebase. Thanks!

0

1

Khue Le

@netw0rkf10w

5 years

@tscholak @stanfordnlp @chrmanning You mean Clark ( @clark_kev ) et al.? There’s a difference ;)

1

0

1

Khue Le

@netw0rkf10w

7 years

#CVPR 2018 received 3303 valid submissions, breaking the record of CVPR 2017's 2680 submissions.

0

1

Khue Le

@netw0rkf10w

5 years

@bermanmaxim @cvpr2020 Congratulations Maxim! It’s time for me to catch up my (former) labmates. @ebelilov @puneetdokania

0

1

Khue Le

@netw0rkf10w

2 years

@kjoshanssen Maybe just simply remove the second sentence? "By Theorem 1, there exists an x with property P(x)."

0

1

Khue Le

@netw0rkf10w

5 months

@zacharynado @aaron_defazio Thanks. What’s the deadline for it to be added to the leaderboard?

1

0

1

Khue Le

@netw0rkf10w

7 years

Thrilled to see a boost in performance of @PyTorch in the upcoming months!

Soumith Chintala

@soumithchintala

7 years

It's GREAT to learn that we are slower, gives us easy room to improve. Thanks a lot to @gneubig and dynet. Details:

1

14

99

0

1

Khue Le

@netw0rkf10w

4 years

@laurent_besacie @coling2020 @Hang_Le @didier_schwab @facebookai Pinging @formiel

0

1

Khue Le

@netw0rkf10w

4 years

Turns out this "professor" doesn't know how to do research.

Steve Hanke

@steve_hanke

4 years

Contrary to an image I posted last week, #Vietnam turns out to have a "perfect" record in its fight against coronavirus. Official Vietnamese data indicate that Vietnam has not suffered any coronavirus fatalities--zero (0) deaths.

119

16

75

0

1

Khue Le

@netw0rkf10w

5 months

@SunnySanyal9 @aaron_defazio @jeankaddour I understand that you cited a 2022 paper for "latest weight averaging" because that's your co-author's paper, but it should be pointed out that this technique has been standard practice since many years. You can find it in, e.g., the Transformer paper (Vaswani et al. 2017).

1

0

1

Khue Le

@netw0rkf10w

6 years

CVPR 2018 papers are available for free:

0

1

Khue Le

@netw0rkf10w

7 years

@drfeifei I love how you care about your students :)

0

1

Khue Le

@netw0rkf10w

4 years

@ducha_aiki @amy_tabb @tdietterich Therefore, they proposed a compromise, while figuring out the best solution. A perfect solution may not exist, but there are ones that are better than others. I proposed one, which I think better than many others (including yours).

1

0

1

Khue Le

@netw0rkf10w

4 years

Woah, such an interesting idea! I feel that this could work, if implemented properly...

Barbara Plank

@barbara_plank

6 years

@ZeerakW @haldaume3 @rktamplayo @kchonyc @emilymbender @zehavoc @thedansimonson @lintool @xtimv @redpony @aria42 yes, you remember right! Thx, @ZeerakW - abandon all deadlines and instead have a rolling one each month -- conferences are then there to present papers as they accumulate. All reviewers in one pot

0

2

9

0

1

Khue Le

@netw0rkf10w

7 years

Videos of CVPR orals and spotlights are available:

0

1

Khue Le

@netw0rkf10w

5 months

@tienhaophung @aaron_defazio Great, thanks. I'll try that in a few days.

0

1

Khue Le

@netw0rkf10w

1 year

Fantastic work! Congrats @TimDarcet and colleagues!

TimDarcet

@TimDarcet

1 year

1/ This week we released DINOv2: a series of general vision encoders pretrained without supervision. Good out-of-the-box performance on a variety of domains, matching or surpassing other publicly available encoders.

5

117

696

0

1

Khue Le

@netw0rkf10w

7 years

Reinforcement Learning: An Introduction, by Richard S. Sutton and Andrew G. Barto Complete draft of the 2nd edition:

0

1