Xin Wang @xinw_ai Twitter profile | Pikagi

Pikagi

Xin Wang

@xinw_ai

3,905

Followers

993

Following

30

Media

233

Statuses

Senior Researcher @MSFTResearch . PhD from @Berkeley_EECS . #artificialintelligence #LLM #multimodal

Seattle, WA

https://t.co/vjQSKjfkUV

Joined August 2013

Don't wanna be here? Send us removal request.

Pinned Tweet

@xinw_ai

Xin Wang

28 days

We are releasing phi-3 mini today! Finally, we have an open-sourced SLM (3.8B) at GPT-3.5 level! Checkout the models and technical report at here 🥳

Tweet card media

microsoft/Phi-3-mini-128k-instruct · Hugging Face

@SebastienBubeck

Sebastien Bubeck

@SebastienBubeck

29 days

phi-3 is here, and it's ... good :-). I made a quick short demo to give you a feel of what phi-3-mini (3.8B) can do. Stay tuned for the open weights release and more announcements tomorrow morning! (And ofc this wouldn't be complete without the usual table of benchmarks!)

40

184

920

2

13

111

Last Seen Profiles

@BAIEbgi

@WalterSantosJr2

@shiioVAL

@_OverCees

@ozwspw

@nadiabaram

@tiedyenation

@brxinbuster

@btcrocksei

@TvSokap

@asolboop

@vkkings007

@element_oh_p

@69jt__

@Avnish261206

@Paffieo

@AgendaQ

@bjorn_k123

@LeMelar

@STYKOcs

@Noodledori1

@Dheeraj41169190

@Octo8gaming

@Phillyconcarne

@hcbsltd

@Sertik4

@GenetixTeam

@7chill_12

@OsoroiCode

@sonodaken

@PalmMiddle

@TrusteeTed

@TheHighwayMen_

@OGDEGENBOYZCLUB

@LincolnLabour

@Jake_NSFW1

@xinw_ai

Xin Wang

3 years

Life update: today is my first day at @MSFTResearch ! Super excited to work with awesome colleagues and friends here in the space of computer vision & machine learning! Reach out to chat. :D

36

12

729

@xinw_ai

Xin Wang

3 years

The computer vision group at @MSFTResearch is looking for research interns! Please retweet to spread the word! 🙏 More at

Tweet media one

6

80

273

@xinw_ai

Xin Wang

3 years

🌀CycConf: Robust Object Detection via Instance-Level Temporal Cycle Confusion at #ICCV2021 . We introduce a new self-supervised task on videos to improve the out-of-domain generalization of object detectors. arXiv: Code:

Tweet media one

2

25

133

@xinw_ai

Xin Wang

2 years

Hi friends at #CVPR2022 , I'm going to talk at the #DynamicNN workshop at 2:30 PM in Great Hall C today. This is my first in-person talk after over two years. Cannot wait to meet you all and chat! Thank the organizers for putting it together! See program details below. 👇

Tweet media one

Tweet media two

Tweet media three

Tweet media four

1

14

135

@xinw_ai

Xin Wang

10 months

Made it to ICML this morning. I will be at Microsoft Booth from 10-11am Tuesday and Wednesday this week. And ping me if you are around too. Looking forward to meeting new and old friends 🤩

Tweet media one

2

2

119

@xinw_ai

Xin Wang

10 months

It turns out **refocusing** the attention of a pre-trained model can help build customized models for your new tasks at test time. TOAST is lightweight (on par or even less than LoRA, etc.) and sees substantial improvements in adapting LLMs, vision foundation models, etc. And

@baifeng_shi

Baifeng

10 months

We have lots of cool algorithms (fine-tuning, LoRA, prompt tuning...) that can adapt a pre-trained model to a downstream task, but what do they MISS? Surprisingly, we find models trained with these algorithms have trouble with focusing their attention!

Tweet media one

8

121

628

2

18

111

@xinw_ai

Xin Wang

1 year

Thanks @_akhaliq for sharing the work! We will release the data and demo shortly. Stay tuned! Great collaboration with @tianjun_zhang @YiZhangZZZ @VibhavVineet @neelsj

@_akhaliq

AK

1 year

Controllable Text-to-Image Generation with GPT-4 introduce Control-GPT to guide the diffusion-based text-to-image pipelines with programmatic sketches generated by GPT-4, enhancing their abilities for instruction following. Control-GPT works by querying GPT-4 to write TikZ code,

Tweet media one

8

63

283

1

14

75

@xinw_ai

Xin Wang

2 years

BDD100K tracking challenge is live at #CVPR2022 Looking forward to your participation!

@DrFisherYu

Fisher Yu

2 years

Our BDD100K () tracking challenges at CVPR 2022 are now live! We hope our challenges can lead to new advances in computer vision and new pragmatic algorithms on self-driving cars. Challenge details:

1

27

124

0

12

65

@xinw_ai

Xin Wang

2 years

New work on OOD detection! 👇👇

@xuefeng_du

Xuefeng Du

2 years

Excited to release our #CVPR2022 oral paper STUD, a powerful unknown-aware object detection framework that safeguards against OOD objects. STUD is the first to leverage videos in the wild and teaches models to tell apart known and unknowns. (1/n) Paper: .

Tweet media one

3

14

91

0

5

52

@xinw_ai

Xin Wang

2 years

#CVPR2022 In today's Oral 3.2.3 (1330) and Poster 3.2 (1430-1700), we will present STUD which distills unknown objects from videos in the wild and regularizes the model’s decision boundary to discriminate the in-distribution and OOD objects.

Tweet media one

1

7

46

@xinw_ai

Xin Wang

6 months

Super excited about the announcement of Phi-2 from our team at MSR!!

@SebastienBubeck

Sebastien Bubeck

@SebastienBubeck

6 months

Microsoft💜Open Source + SLMs!!!!! We're so excited to announce our new *phi-2* model that was just revealed at #MSIgnite by @satyanadella ! At 2.7B size, phi-2 is much more robust than phi-1.5 and reasoning capabilities are greatly improved too. Perfect model to be fine-tuned!

Tweet media one

Tweet media two

18

87

489

0

1

48

@xinw_ai

Xin Wang

1 year

The human in the loop learning workshop at #NeurIPS2022 is happening at room 396. We have a great lineup of great speakers! @PeterStone_TX is giving a talk now on human in the loop learning for robot navigation and task learning from implicit human feedback.

Tweet media one

1

5

46

@xinw_ai

Xin Wang

2 years

Hi ML friends, I'm flying to Baltimore today for #ICML2022 . Feel free to email or DM me to chat! @baifeng_shi and I will present "Visual Attention Emerges from Recurrent Sparse Reconstruction" on Thursday, July 21 in Hall F. Come over to chat with us!

Tweet media one

2

2

43

@xinw_ai

Xin Wang

8 months

You can now use Phi-1 and Phi-1.5 models on Huggingface and Azure ML!👇

@SebastienBubeck

Sebastien Bubeck

@SebastienBubeck

8 months

phi-1.5 & phi-1 are available right now on @huggingface & @Azure ML! We can't wait to see what the community will discover with them. The phi-1.5 team Yuanzhi Li @EldanRonen @allie_adg @suriyagnskr is ready to answer questions too!

2

19

129

1

7

41

@xinw_ai

Xin Wang

3 years

Consider applying if you are interested! Feel free to contact me if you are interested in the CV/ML area. 😀

@besanushi

Besmira Nushi 💙💛

3 years

The MSR Undergraduate Research Internship program is now accepting applications This is truly one of a kind opportunity for undergrads to get early experience and mentorship in research. Deadline November 19th.

4

39

106

2

3

40

@xinw_ai

Xin Wang

3 years

Check out our new ICCV 21' work 👇 👇👇We introduce a new benchmark with a set of metrics for online continual object detection. The data spans a 9-month lifetime of a graduate student, Krishna, and provides a realistic setting to learn through a person's eyes. Enjoy! 🚶🚶‍♂️🚶‍♀️

@wang_jianren

Jianren Wang

3 years

Wanderlust: Online Continual Object Detection in the Real World at #ICCV2021 with @xinw_ai , Yue and Abhinav We present a new online continual object detection benchmark with an egocentric video dataset. Paper: Code:

3

2

18

1

2

40

@xinw_ai

Xin Wang

8 months

I am excited to announce our HoloAssist work at #ICCV ! Get your hands on the dataset at , released under a permissive license. Meet my collaborator @big_stamp at the poster session this Friday from 02:30-04:30 PM. Huge thanks to the coauthors and

@MSFTResearch

Microsoft Research

8 months

HoloAssist is a new multimodal dataset consisting of 166 hours of interactive task executions with 222 participants. Discover how it offers invaluable data to advance the capabilities of next-gen AI copilots for real-world tasks:

Tweet media one

1

10

40

2

5

38

@xinw_ai

Xin Wang

4 years

The paper was accepted at #icml2020 ! Looking forward to seeing you virtually in July!

@xinw_ai

Xin Wang

4 years

We find that a simple two-stage fine-tuning approach can outperform the previous meta-learning methods by a large margin on few-shot object detection. New results on Pascal VOC, COCO and LVIS! w/ @thomasehhh @DrFisherYu arXiv code

Tweet media one

0

6

17

2

2

35

@xinw_ai

Xin Wang

2 years

Excited to announce the L3D-IVU workshop at #CVPR2022 . Please check out the details below. Shout-out to @mennatullahSiam and the organizing team for putting the event together!

@CSProfKGD

Kosta Derpanis

2 years

We are organizing the #CVPR2022 Workshop on Learning with Limited Labelled Data. Check out the invited speakers and call for papers! @mennatullahSiam @xinw_ai @dvazquezcv @prlz77 @ILaradji @GabrielHuang9 @BOreshkin @hugo_larochelle @dimadamen

Tweet media one

Tweet media two

Tweet media three

2

21

120

1

0

34

@xinw_ai

Xin Wang

2 years

PSI + HoloLens is coming! A great platform supporting interactive AI research!

@MSFTResearch

Microsoft Research

2 years

Platform for Situated Intelligence is a framework for building AI systems that can perceive and act in the world in real time, and researchers have recently extended it to support mixed-reality applications targeting the HoloLens 2!

1

16

60

0

3

32

@xinw_ai

Xin Wang

8 months

I am glad to give a tutorial talk on data-efficient learning with top-down attention remotely. Thanks @SifeiL and @xiaolonw for the invitation. Looking forward to seeing you all online. :)

1

1

31

@xinw_ai

Xin Wang

2 years

It was great to see so many people have come to visit our poster. Great work @baifeng_shi !

Tweet media one

Tweet media two

@xinw_ai

Xin Wang

2 years

Hi ML friends, I'm flying to Baltimore today for #ICML2022 . Feel free to email or DM me to chat! @baifeng_shi and I will present "Visual Attention Emerges from Recurrent Sparse Reconstruction" on Thursday, July 21 in Hall F. Come over to chat with us!

Tweet media one

2

2

43

0

1

25

@xinw_ai

Xin Wang

10 months

@DynamicWebPaige Thanks for featuring our work! Yeah, we are actively looking into upgrading Gorilla with LLama2. Stay tuned! 😉

2

1

23

@xinw_ai

Xin Wang

2 years

@WenhuChen I got Covid after attending CVPR. I feel like parties (even outdoors) and eating together might be the source of infection. Most people wear masks during sessions. Tho I didn’t regret attending CVPR, I indeed felt very sick - fever for 3 nights and bad cough. So take care!

4

0

23

@xinw_ai

Xin Wang

2 years

It was a lot of fun at the workshop today. Thank everyone for joining us. See you next time! 😬😬😬

@CSProfKGD

Kosta Derpanis

2 years

And that’s a wrap for the #CVPR2022 Workshop on Learning with Limited Labelled Data for Image and Video Understanding. HUGE thank you to all the participants and our sponsors!

Tweet media one

Tweet media two

Tweet media three

4

9

55

1

1

20

@xinw_ai

Xin Wang

2 years

Our workshop is happening in 7 hours! The workshop will be located in Room 224 starting from 8:45 AM New Orleans time. Our speakers will also join the panel discussion at 4:45PM moderated by @CSProfKGD and me. Enjoy! 😃😃😃

@CSProfKGD

Kosta Derpanis

2 years

We are just five days out from the #CVPR2022 Workshop on Learning with Limited Labelled Data for Image and Video Understanding ( #L3DIVU2022 ). Check out the schedule; includes 7 OUTSTANDING invited speakers! Full details: 20 June 2022, Room 224

Tweet media one

1

17

64

0

4

17

@xinw_ai

Xin Wang

4 years

We find that a simple two-stage fine-tuning approach can outperform the previous meta-learning methods by a large margin on few-shot object detection. New results on Pascal VOC, COCO and LVIS! w/ @thomasehhh @DrFisherYu arXiv code

Tweet media one

0

6

17

@xinw_ai

Xin Wang

4 years

The Human in the loop Learning (HILL) workshop @icmlconf submission deadline is now extended to **June 20**! Please consider submitting if you are interested! See for more information! w/ @LotusSapphire @DrFisherYu @jiajunwu_cs Trevor Darrell #icml2020

Tweet media one

Tweet media two

0

3

17

@xinw_ai

Xin Wang

3 years

Congrats @mejoeyg ! Fantastic news and well deserved! It has been awesome working with Joey in the past six years. I cannot wait to celebrate sometime in person with the group! :D

@profjoeyg

Joey Gonzalez

3 years

I am very excited to announce that my research group just received tenure @UCBerkeley . Alright, technically I received tenure, but I could not have done this without the hard work of my amazing team of students and colleagues. Queue awards music (1/4)

40

10

540

0

0

15

@xinw_ai

Xin Wang

4 years

We are hosting a live QA for our #CVPR2020 paper at 9-11am PDT and 9-11PM PDT today! We propose ShapeProp, which learns saliency propagation through message passing for semi-supervised instance segmentation. Paper Code:

0

3

13

@xinw_ai

Xin Wang

5 months

Excited to share the latest progress of phi-2 from our group! Check out the blog post for more details! 👇👇👇

@SebastienBubeck

Sebastien Bubeck

@SebastienBubeck

5 months

Phi-2 numbers, finally! We're seeing a consistent ranking: phi-2 outperforms Mistral 7B & Gemini Nano 2* (*on their reported benchmarks) and is roughly comparable to Llama 2-70B (sometimes better, sometimes worse). Beyond benchmarks, playing with the models tells a similar story.

Tweet media one

21

83

587

0

0

14

@xinw_ai

Xin Wang

3 years

CycConf will be presented in Session 7, Oct 13 & Oct 15. Look forward to seeing you there.

@xinw_ai

Xin Wang

3 years

🌀CycConf: Robust Object Detection via Instance-Level Temporal Cycle Confusion at #ICCV2021 . We introduce a new self-supervised task on videos to improve the out-of-domain generalization of object detectors. arXiv: Code:

Tweet media one

2

25

133

0

0

12

@xinw_ai

Xin Wang

3 years

Check out our new work! 👇👇👇

@_amirbar

Amir Bar

3 years

Excited to share DETREg, our new work on unsupervised pretraining for object detection with transformers. Compared to previous works, the key idea in DETReg is attempting to learn object detection in the unsupervised pre-training stage. @berkeley_ai , @TelAvivUni , @NVIDIAAI

Tweet media one

1

8

51

0

0

12

@xinw_ai

Xin Wang

2 years

Great opportunity for graphics researchers and students! The mentorship program looks exciting! :)

@wigraphorg

WiGRAPH

2 years

We’ve decided to extend the deadline for #WiGRAPHRisingStars until April 8, 11:59 pm Anywhere on Earth. Last week to apply!!! Don’t miss out on this great opportunity to receive mentorship, especially career and job-search advice, from top researchers in industry and academia!

Tweet media one

2

14

35

0

1

12

@xinw_ai

Xin Wang

3 years

Interesting findings!

@clmich

Claudio Michaelis

4 years

More classes is all you need for one-shot object detection. It turns out we can almost close the gap between the detection of known and novel objects simply by increasing the number of categories. @clmich @alxecker @MatthiasBethge [1/7]

Tweet media one

1

20

90

0

1

10

@xinw_ai

Xin Wang

3 years

The topics cover but are not limited to representation learning, few-shot/continual learning, domain generalization, embodied AI, and beyond. More at our group page @neelsj @yalesong @VibhavVineet

Computer Vision Group - Microsoft Research

The Computer Vision Group conducts research in areas spanning AI, Machine Learning, Computer Vision, and Computational Photography. Our mission is to conduct state of the art foundational research...

www.microsoft.com

1

1

8

@xinw_ai

Xin Wang

3 years

Please contact us at msrcvinternapply @microsoft .com if you are interested.

3

0

8

@xinw_ai

Xin Wang

10 months

Congrats to both the Gorilla team and the Vicuna team! 😀

@profjoeyg

Joey Gonzalez

10 months

I am excited to announce that two of the LLMs from my group (Gorilla and Vicuna) are on AI Business’s top 12 models. Congrats @shishirpatil_ , @tianjun_zhang , @xinw_ai , and the @lmsysorg team. We are looking forward to working with @Meta on Llama-v2 versions.

2

7

51

0

0

8

@xinw_ai

Xin Wang

2 years

A few of us ( @_amirbar for sure) will be around at the poster. Please drop by and chat about self-supervised pretraining for object detection. :D

@_amirbar

Amir Bar

2 years

If you are at CVPR, come visit our poster today! Thursday 2:30-5:00, poster 127.

1

6

43

0

0

7

@xinw_ai

Xin Wang

3 years

@MSFTResearch Thanks! Look forward to the new adventure! 🤩

0

0

5

@xinw_ai

Xin Wang

3 years

@FrancisYan_ Great initiative, Francis! I would be happy to join you and chat/give feedback on the proposals. :D

1

0

4

@xinw_ai

Xin Wang

3 years

@TrungBi72462511 @MSFTResearch Hi, actually yes. Here is an undergraduate research program at MSR.

1

1

4

@xinw_ai

Xin Wang

10 months

@DynamicWebPaige Thanks for organizing the event! It was a lot of fun!

1

0

3

@xinw_ai

Xin Wang

1 year

The afternoon session just started. Dr. @danbohus from Microsoft Research is giving a talk on mixed reality human in the loop learning!

Tweet media one

1

0

3

@xinw_ai

Xin Wang

8 months

@big_stamp @big_stamp has made a detailed thread on the work! Check it out!

@taeinkwon1

Taein Kwon

8 months

📢Are you interested in how interactive AI assistants can collaborate with humans in the real world? If so, you don’t want to miss our #ICCV2023 paper, “HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World”!

1

1

9

0

0

3

@xinw_ai

Xin Wang

3 years

Instead of finding the nearest neighbors, we encourage the model to explore the latent structures across different instances by purposely assigning higher matching scores for the more different objects. As shown below, CycConf maximizes the entropy of the matching scores.

Tweet media one

1

0

3

@xinw_ai

Xin Wang

3 years

@dchaplot @rsalakhu Congrats!!

0

0

3

@xinw_ai

Xin Wang

1 year

Our last invited talk is from Carlos Quintero-Peña from Rice University on motion planning. After this talk, we will be hosting a panel discussion from 4-5pm with Profs @samikaski @CynthiaRudin , Dr. @IMordatch at Room 396 and Prof. @AlisonGopnik and Dr. @danbohus in zoom.

Tweet media one

1

0

3

@xinw_ai

Xin Wang

1 year

Gorilla's another milestone achieved :)

@shishirpatil_

Shishir Patil

1 year

🦍Big news! Gorilla is now Apache 2.0 licensed🤩We are delighted to welcome 2 new models into the Gorilla family ⛳️Use commercially with zero obligations! Our colab has been used for 6000+ invocations in the last week 🚀Check it out:

4

23

101

0

0

3

@xinw_ai

Xin Wang

10 months

@lmsysorg Great work and lightning speed too! 👍

0

0

3

@xinw_ai

Xin Wang

1 year

Prof @samikaski is giving a talk now on collaborative AI for assisting virtual laboratories.

Tweet media one

1

0

3

@xinw_ai

Xin Wang

1 year

We just concluded the talks of the day and the poster session is still going on in Room 396. Thanks our speakers and panelists for the great talks and discussion. Also congratulations to the best paper and runner-ups recipients!

Tweet media one

Tweet media two

Tweet media three

Tweet media four

0

0

3

@xinw_ai

Xin Wang

2 years

@sainingxie @NYU_Courant @nyuniversity @CILVRatNYU Congrats, Saining!!

0

0

3

@xinw_ai

Xin Wang

1 year

Prof. @AlisonGopnik from Berkeley is giving a talk on what children can do while large models cannot.

Tweet media one

1

0

3

@xinw_ai

Xin Wang

3 years

Inspired by prior works on self-supervised learning in time (e.g, TimeCycle by @xiaolonw , Allan Jabri et al., TCC by Dwibedi et al.), we form a time cycle for self-supervision. The key difference is how to identify the **soft target** in the adjacent frame.

1

0

3

@xinw_ai

Xin Wang

10 months

@vmuaddib And for your second question, TOAST applies to open-domain tasks. For example, the attention can be re-steered to the most relevant features instead of the entire input sequence.

1

0

2

@xinw_ai

Xin Wang

10 months

@vmuaddib Thanks for your interest in our work! Let me break down your questions. Yes, you are right that the doubled forward passes in TOAST is a double-edged sword, which can boost accuracy but may sacrifice the compute efficiency.

1

0

2

@xinw_ai

Xin Wang

2 years

We also compare the attention map of VARS and self-attention with the human eye fixation map, showing that VARS has a more consistent behavior to human attention.

Tweet media one

1

0

2

@xinw_ai

Xin Wang

10 months

@YangYou1991 I’m flying to Honolulu tonight. Let’s catch up there 🤩

1

0

2

@xinw_ai

Xin Wang

3 years

@boysdontcrei @MSFTResearch Thank you! I will definitely share my experience at MSR. Looking forward to the new adventure!

0

0

2

@xinw_ai

Xin Wang

3 years

@zhoubolei @UCLAengineering @CS_UCLA congratulations!!!

0

0

2

@xinw_ai

Xin Wang

3 years

@danfei_xu @gtcomputing @ICatGT @GTrobotics @mlatgt Congrats! Well deserved! Are you gonna spend your gap year in the Seattle office or the Bay Area? Let me know if you are around. 🎉 🎉🎉

1

0

1

@xinw_ai

Xin Wang

3 years

@Kaushik171 @TaliaRinger @MSFTResearch I just followed you. You should be able to DM me now. :)

0

0

2

@xinw_ai

Xin Wang

3 years

Please check out the paper if you are interested. The code and model checkpoints are released at the link above. 🤗🤗🤗

1

0

2

@xinw_ai

Xin Wang

3 years

We find this regularization helps the model generalize to unseen domains at test time. We also construct an out-of-domain generalization benchmark on BDD100K videos and Waymo open data and show the effectiveness of the method across different times of day, camera angles, etc.

1

0

2

@xinw_ai

Xin Wang

3 years

@Kaushik171 @TaliaRinger @MSFTResearch Sure, happy to help!

1

0

2

@xinw_ai

Xin Wang

1 year

Prof @CynthiaRudin from Duke University is giving a talk on Let’s Give Domain Experts a Choice by Creating Many Approximately-Optimal Machine Learning Models

Tweet media one

1

0

2

@xinw_ai

Xin Wang

1 year

@timshi_ai Thanks for sharing the work, Tim! ☺️

1

0

2

@xinw_ai

Xin Wang

3 years

@SaynaEbrahimi @UCBerkeley @GoogleAI Congrats!!

1

0

2

@xinw_ai

Xin Wang

2 years

This approach is based on our ICCV 2021 work: Cycle-Confusion to improve OOD generalization of object detectors. See this thread 👇

@xinw_ai

Xin Wang

3 years

🌀CycConf: Robust Object Detection via Instance-Level Temporal Cycle Confusion at #ICCV2021 . We introduce a new self-supervised task on videos to improve the out-of-domain generalization of object detectors. arXiv: Code:

Tweet media one

2

25

133

1

0

0

@xinw_ai

Xin Wang

2 years

@xxxnell Ah I just opened my DM. Didn’t know it was closed the whole time. I believe my collaborator @baifeng_shi has emailed you and we can find a time to chat!

0

0

2

@xinw_ai

Xin Wang

2 years

Find more details in the paper at Code at

Tweet card media

GitHub - bfshi/VARS: Official code for `Visual Attention Emerges from Recurrent Sparse Reconstruc...

Official code for `Visual Attention Emerges from Recurrent Sparse Reconstruction' (ICML 2022) - bfshi/VARS

0

0

2

@xinw_ai

Xin Wang

2 years

We then show that visual attention **emerges** naturally if we iteratively solve the sparse reconstruction problem. The VARS module (shown 👇) achieves visual attention by adding sparse reconstruction blocks into feedforward networks.

Tweet media one

1

0

1

@xinw_ai

Xin Wang

8 years

Excited about the coming #NIPS2015 Look forward to meeting all the top researchers in the field!

0

0

0

@xinw_ai

Xin Wang

2 years

We show that it is equivalent to optimizing a sparse reconstruction (SR) problem - adding additional SR blocks which solve the sparse reconstruction of the output from the previous block, using a fixed dictionary learned from data.

Tweet media one

1

0

1

@xinw_ai

Xin Wang

2 years

STUD is a collaboration with @xuefeng_du and @SharonYixuanLi from UW-Madison. Paper: Code:

GitHub - deeplearning-wisc/stud: source code for CVPR'22 paper "Unknown-Aware Object Detection:...

source code for CVPR'22 paper "Unknown-Aware Object Detection: Learning What You Don’t Know from Videos in the Wild" - deeplearning-wisc/stud

1

0

1

@xinw_ai

Xin Wang

3 years

@ajayj_ Yeah it is.🤣

0

0

1

@xinw_ai

Xin Wang

10 months

@vmuaddib To deal with it, we can treat the first pass as a coarse processing pass that can use a sparse network or use input with lower resolution to mitigate the computational cost.

1

0

1

@xinw_ai

Xin Wang

11 months

@polynoamial @OpenAI Congrats, Noam!!!

0

0

1

@xinw_ai

Xin Wang

3 years

@chim_mezie @bexxmodd @MSFTResearch Thank you!

0

0

1

@xinw_ai

Xin Wang

3 years

@Diyi_Yang @ICatGT @gtcomputing @mlatgt Congrats Diyi!

0

0

1

@xinw_ai

Xin Wang

1 year

@SharonYixuanLi Sorry to hear that. Take care!!!

0

0

1

@xinw_ai

Xin Wang

1 year

@samikaski @FCAI_fi @OfficialUoM Thank you for the insightful talk today. Looking forward to more discussion in the panel in the afternoon!

0

0

1

@xinw_ai

Xin Wang

10 months

@vmuaddib @baifeng_shi and I have been thinking about this direction too. We have some promising results on Alpaca in the paper but definitely worth further investigation on this topic. Happy to chat more if you are interested.

1

0

1

@xinw_ai

Xin Wang

3 years

@ylzou_Zack I didn't observe improvements on in-domain data points. The performance remains the same compared to a vanilla Faster R-CNN detector.

0

0

1

@xinw_ai

Xin Wang

2 years

@SebastienBubeck @geoishard Congratulations, Seb. This is so cool!

0

0

1

@xinw_ai

Xin Wang

1 year

@IMordatch from Google Brain is giving a talk on language models and interactive decision making.

Tweet media one

1

0

1

@xinw_ai

Xin Wang

11 months

@yalesong Wonderful show! 🥳🥳

0

0

1

@xinw_ai

Xin Wang

2 years

@moElhoseiny Thanks for the discussion today. It was a home run! :D

0

0

1

@xinw_ai

Xin Wang

3 years

@FrancisYan_ @MSFTResearch bitter and sweet experience 😜

0

0

1

@xinw_ai

Xin Wang

3 years

@NeerajaJY @CockrellSchool @UTAustin @utexasece Congrats, Neeraja!!!

1

0

1

@xinw_ai

Xin Wang

2 years

@TianxingH Congrats 🎊🎉🍾🎈

0

0

1

@xinw_ai

Xin Wang

3 years

@Saikallis9012 @MSFTResearch Thank you!!

0

0

1

@xinw_ai

Xin Wang

3 years

@FrancisYan_ wow!! I wish we can have an in-person commencement as well. 😔

0

0

1

@xinw_ai

Xin Wang

10 months

@ying11231 Congrats！🥳🥳

1

0

1

@xinw_ai

Xin Wang

3 years

@kexinrong @gatech_scs Congrats Kexin! 🎉🎉🎉

0

0

1

@xinw_ai

Xin Wang

2 years

VARS can be plugged into NNs just like Self-Attention, a key component in the renowned vision transformers. Actually, we can show mathematically that Self-Attention is a special case of VARS --- a loose optimization of the ODE formulation without sparsity constraints. 👇

Tweet media one

1

0

1

@xinw_ai

Xin Wang

8 months

@SharonYixuanLi @techreview Congrats, Sharon!! Well deserved :)

0

0

1

@xinw_ai

Xin Wang

3 years

@gazebofiftytry @MSFTResearch Thanks! Go bears!

0

0

1

@xinw_ai

Xin Wang

6 months

@realSharonZhou Wow! Didn’t know you were around. Let’s hang out if you are still here 🤩

1

0

1