Dongyoon Han

Research scientist @NAVER AI Lab. Adjunct professor @KAIST GSAI

[Dongyoon Han @karusun1]

@JungWooHa2 invitation 꽤 빨리 오던데 아직 더 부족한가 보네요.

[Junyoung Seo @jyseo_cv]

What if a world model could render not an imagined place, but the actual city?

We introduce Seoul World Model, the first world simulation model grounded in a real-world metropolis.

TL;DR: We made a world model RAG over millions of street-views.

proj: https://t.co/Bx4KUAqrRs

[Lucas Beyer (bl16) @giffmana]

soooo... how many papers do we think are invalidated by this? And now think about how many other bugs there must be in any re-implementations of... basically anything.

[Sangdoo Yun @oodgnas]

Paper: https://t.co/sjnekBvIxs
Code: https://t.co/kKtn2DadJc
Project page: https://t.co/aw9ghziXZx

@JangHyun_k, @karusun1, Sangdoo Yun

[Sharon Li @SharonYixuanLi]

Our work on "DaWin: Training-Free Dynamic Weight Interpolation for Robust Adaptation" has been accepted to #ICLR2025.

Kudos to @Changdae_Oh for driving this nice effort!

Paper: https://t.co/vMWGADq667

[Changdae Oh @Changdae_Oh]

Weight interpolation (more generally, model merging) method is recently becoming a de facto standard used in frontier-level AI, e.g., Llama 3.2🦙

[1/8] Check the latest achievement by our new method, DaWin, which will be presented at AFM@NeurIPS'24🥳
- 📃:https://t.co/N6CnvbSZy1

[Sangdoo Yun @oodgnas]

Do we really need multiple models for weight averaging (e.g., model soups)? Our answer: just two! We introduce Model Stock and share interesting findings on fine-tuned weights. #ECCV2024. Don’t miss our poster at AM 10:30 - 12:30, #110.

[Sangdoo Yun @oodgnas]

If you're interested in position encoding, especially for vision tasks, don’t miss our ECCV poster presentation!
Paper: https://t.co/QVTtK695T5 #ECCV2024
Poster: 4:30 PM - 6:30 PM at #165
@bhheo90 Song Park @karusun1

[Dongyoon Han @karusun1]

Dense connections, originally introduced in DenseNets finally came back! Visit our ECCV poster (#335, 16:30-18:30) if you're interested in how they have revived!!!
#ECCV2024

Paper: https://t.co/MLUp8K8vlf

[Dongyoon Han @karusun1]

@ramealexandre + I've seen model-stocked models already working with OpenChat3.5. Actually, we're exploring multiple fields, including NLP, to ensure our method's compatibility with LLMs. We expect there would be a more effective way.

[Dongyoon Han @karusun1]

@ramealexandre Thanks for sharing the swiftly implemented repository. I haven't checked the code's validity, but our simple method should ensure no problems with implementation (we'll release our code soon after an internal review). Sorry for the delay.

[Seong Joon Oh @coallaoh]

Download annotation byproducts for ImageNet & COCO.

ImageNet-AB https://t.co/qqkgvWwvIQ

COCO-AB:
https://t.co/4nugJhWt47

@karusun1 @junsukchoe @john_jyc @minsuk_chang @oodgnas @jeanysong

[Seong Joon Oh @coallaoh]

1/9 🧵 Looking forward to presenting our work

Neglected Free Lunch -- Learning Image Classifiers Using Annotation Byproducts

at #ICCV2023 in October. Here's a quick overview. 👇 (links at the end)

[Jim Fan @DrJimFan]

Kaiming He, inventor of ResNet, is leaving industry to join MIT faculty in 2024!! He’s one of the most impactful figures in deep learning.

- Residual layer is a fundamental building block of LLMs.
- Faster/Mask R-CNN are industrial standards for image segmentation and robot perception stack.
- Panoptic segmentation redefined a research sub-field in vision.
- Mask AutoEncoder (MAE) is among the best general-purpose self-supervised algorithms for computer vision and beyond.
- Before MAE, Momentum Contrast (MoCo) was a SOTA contrastive learning technique.
- SlowFast network was among the default backbones for video learning until ViTs took over.
- Too many other groundbreaking works to enumerate …

I recently observe an exodus of researchers from big techs to academia. It’s an interesting movement given the current LLM gold rush 🤔

Normally we congratulate someone who joins MIT, but this time I congratulate @MITEECS to have Kaiming! 🎉

[SanghyukChun @SanghyukChun]

SeiT will appear at ICCV 2023! So excited to visit Paris 😊

[Sangdoo Yun @oodgnas]

Excited to introduce our new paper, 'Augmenting Sub-model to Improve Main Model' (AugSub)! 🚀

Paper: https://t.co/4SngnhSCXJ
Code: https://t.co/PhK7neun1z

We've been tackling an ongoing problem -- conducting strong regularization *without* harming loss convergence. 🔍 [1/3]