🚨We uncover a new vulnerability- Pre-Fine-Tuning Weight Recovery With a few LoRA fine-tuned models we recover the pre-fine-tuning weights🏋️of SoTA models, undoing Stable Diffusion personalization training and Mistral alignment😈 Project: vision.huji.ac.il/spectral_detun… 🧵👇

Check out our latest work! The tl:dr is that finetuning can be reversed in some cases, which presents a security threat. This is particularly serious when using finetuning for model alignment or safety training.

🚨New Dataset Distillation paper 🚨 We propose PoDD, a new dataset distillation setting for a tiny🤏under 1 image-per-class (IPC) budget❗In this CIFAR100 example, current SoTA is 35.5% acc using ~100k pixels, PoDD gets 35.7% with ~40k pixels Project: vision.huji.ac.il/podd 🧵👇

My first paper 🥳 We propose PoDD, a new dataset distillation setting for a tiny🤏under 1 image-per-class (IPC) budget❗Getting SoTA performance with as little as 30% of the pixels compared to previous methods 🚀 Eliahu Horwitz Yedid Hoshen Project: vision.huji.ac.il/podd 🧵👇

In our work led by Benjamin Feuer, TuneTables, we scale TabPFN to larger datasets by distilling large sets of labelled datasets into a much shorter soft prompt. Paper: arxiv.org/abs/2402.11137 Notebook: kaggle.com/code/benfeuer/… 🧵🧵🧵🧵🧵

Distilling Datasets Into Less Than One Image by Asaf Shul Eliahu Horwitz Yedid Hoshen #AI #MachineLearning #DeepLearning

Thank you AK for sharing! Check out our project webpage: objectdrop.github.io And congrats to my collaborators Daniel Winter Matan Cohen Shlomi Fruchter Yael Pritch Alex Rav-Acha

We introduce ObjectDrop, our recent Google AI project, aimed at achieving photorealistic object removal and insertion. Explore our project page: objectdrop.github.io Arxiv: arxiv.org/abs/2403.18818

1/6 ObjectDrop is a supervised method centered on a counterfactual dataset for object removal and insertion. By combining both models, we can seamlessly move objects within an image.

Some insane work by colleagues at Google. Really impressive

My team with Yuv Lem won🥇first place in the defense track of the Capture The Flag-LLM Challenge at SaTML Conference. In the competition, each team hides a secret in an LLM context, and other teams try to discover it. Here's how we used ideas from Judo and Cooking to do it👇 🧵1/11

🚨New paper On the Origin of Models🚨 We define the Model Tree🌳 to describe the hereditary relations between models. We propose the task “Model Tree Heritage Recovery” (MoTHer Recovery) for discovering Model Trees via model weights Project: vision.huji.ac.il/mother 🧵👇

Check out our latest work!

🚨New Paper on Deepfake Detection🚨 We introduce LaDeDa: a state-of-the-art patch-based deepfake detector utilizing only 9x9 patch information❗ We also present Tiny-LaDeDa, a fast, compact, yet accurate distilled version for edge devices. Project: vision.huji.ac.il/ladeda/ 👇

I am excited to share my first work: "Dataset Size Recovery from LoRA Weights". Ever wondered if you could find out how many samples was a model trained on using just its weights? Well now you can! Project: vision.huji.ac.il/dsire/ 👇

ObjectDrop is accepted to #ECCV2024! 🥳 In this work from Google AI we tackle photorealistic object removal and insertion. Congrats to the team: Matan Cohen Shlomi Fruchter Yael Pritch  Alex Rav-Acha Yedid Hoshen Checkout our project page: objectdrop.github.io

ObjectDrop accepted to ECCV! 🥳🎉

With friends at Google we announce 💜 Magic Insert 💜 - a generative AI method that allows you to drag-and-drop a subject into an image with a vastly different style achieving a style-harmonized and realistic insertion of the subject (Thread 🧵) web: magicinsert.github.io

Another chance to see our poster for those that missed ICML

Amazing work by my colleagues!