Donut 🍩 : Document Understanding Transformer without OCR
— AK (@ak92501) December 1, 2021
abs: https://t.co/A644UXgUuG
achieves sota performance on various document understanding tasks in public benchmark datasets and private industrial service datasets pic.twitter.com/2broMiK9r5
Pyramid Adversarial Training Improves ViT Performance
— AK (@ak92501) December 1, 2021
abs: https://t.co/oaxB6Q99R2
new sota for ImageNet-C (41.4 mCE), ImageNetR (53.92%), and ImageNet-Sketch (41.04%) without extra
data, using only the ViT-B/16 backbone and pyramid
adversarial training pic.twitter.com/afdSG0J35i
One kinda nice CLI habit I've picked up recently is when I need to go check something or run something else in the middle of composing a command, I just stick a "#" at the beginning and hit enter. Then it's in my history and I can go back to it later.
— Brendan Dolan-Gavitt (@moyix) December 1, 2021
Differential Inference: A Criminally Underused Tool (https://t.co/zSwhvk806r)
— Sasha Rush (@srush_nlp) November 30, 2021
An annotated talk about elementary probability (coins&dice) in pytorch. Nothing new, just think we should mostly do discrete inference with auto-diff.
Slides: https://t.co/tJHQlLwNrp pic.twitter.com/rq37IuEuh6
Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity
— AK (@ak92501) November 30, 2021
abs: https://t.co/U4I6gnNvej
Sparse DETR achieves better performance than Deformable DETR even with only 10% encoder tokens on the COCO dataset pic.twitter.com/gcERMMxUu4
HittER: Hierarchical Transformers for Knowledge Graph Embeddings
— AK (@ak92501) November 26, 2021
abs: https://t.co/TYTZOxVtiB
github: https://t.co/ymOIjKSH1b pic.twitter.com/2I8lWDj4r7
Self-slimmed Vision Transformer
— AK (@ak92501) November 25, 2021
abs: https://t.co/BHQIZVWZlN pic.twitter.com/ToB9DODFwU
I have been thinking a lot about this paper about why explainable AI is not ready for patient-level explainability of AI in healthcare: https://t.co/z9FG2jXe7P
— Christoph Molnar (@ChristophMolnar) November 25, 2021
While I would not go as far as calling it a "false hope", there is food for thought in there.
1/n
Small tip for correctly formatting a collection of Python Notebooks, install @marcwouts 's #jupytext and #black, then run: pic.twitter.com/v4vBQnO3ws
— Martin Renou (@martinRenou) November 23, 2021
You use a CV to prove your worth.
— Radek Osmulski (@radekosmulski) November 22, 2021
But it is a lousy way of doing so.
Your CV competes against many other CVs.
In the end, it is just a stream of words that carry little weight.
It's a tough spot to put yourself in.
How is blogging different?
Combined Scaling for Zero-shot Transfer Learning
— hardmaru (@hardmaru) November 22, 2021
Another data point for Sutton's “Bitter Lesson”: more data, bigger model, bigger batch sizes combined leads to big performance gains for zero-shot learning.
New paper by @hieupham789 and others @GoogleAIhttps://t.co/q97vHUsM45 pic.twitter.com/aXqKb5NaHu
We're using https://t.co/GtrlycEDlf and https://t.co/j99uBsAKg2 which are fantastic tools, but there are significant challenges in adapting these tools to account for interference and seasonality in marketplace outcomes.
— Sean J. Taylor (@seanjtaylor) November 20, 2021
