Excited to be at #CVPR2025 in Nashville! We're presenting our paper, "DiTASK: Multi-Task Fine-Tuning with Diffeomorphic Transformations." 🧵(1/n). Come chat with us at our poster session: .🗓️ Sun, Jun 15 ⏰ 10:30 AM - 12:30 PM CDT .📍 ExHall D, Poster #399

🔗 Project Page: 📰 Paper: 🧑‍💻 Code: (n/n).

We're here @CVPR all week and would love to meet! Let's talk about #MTL, #PEFT, #VisionTransformers, and diffeomorphisms. Reply here to connect with me and the team (@ChaimBaskin and Moshe Eliasof)!. #AI #ComputerVision #MachineLearning #DeepLearning #CVPR2025 #Diffeomorphisms.

The Results: DiTASK sets a new S.O.T.A. with 75% fewer parameters! .✅ +3.22% avg. improvement on PASCAL MTL .✅ >2x improvement on NYUDv2 tasks .✅ 22% faster inference than MTLORA .✅ The performance boost gets even bigger with larger models!

Our solution: 𝐃𝐢𝐓𝐀𝐒𝐊! We introduce a new way to fine-tune that preserves the core visual knowledge in pre-trained models. We freeze the singular vectors (feature directions) and only adapt the singular values (feature importance) using efficient, learnable diffeomorphisms.

𝐓𝐡𝐞 𝐏𝐫𝐨𝐛𝐥𝐞𝐦: Fine-tuning large models for Multi-Task Learning (MTL) is hard. Methods like LoRA force different tasks to compete within a limited subspace, which can hurt performance. This "task interference" is a major hurdle.

Excited to announce I’m joining @INSIGHTLabBGU w/ @ChaimBaskin at BGU’s School of ECE this summer! I’ll be working on ODE-inspired methods to advance representation learning! Huge thanks to Prof. Baskin for making this possible!.

RT @beabevi_: 🚨 New paper at #ICLR2025!. 🧵Introducing Holographic Node Representations, pretrained general-purpose node embeddings that ada….

This is the sign I needed today! 🤡

2025 is also Silver Jubilee of my life 🤪.

It just hit me - This Valentine's day marks Silver Jubilee of no gf. I deserve a commemorative plaque and a ceremony. 🤡😭🎊 💝.

RT @brunofmr: Slides of my presentation "Mathematical Foundations of Graph Foundation Models" yesterday at the AMS Session of the #JMM2025.….

Thrilled to present DiGRAF at @NeurIPSConf with @brunofmr and Moshe Eliasof. Big thanks to everyone who stopped by, asked questions, and offered feedback. DM to grab a coffee and discuss ideas. We’re excited about the directions this work will take—stay tuned for more to come!

RT @beabevi_: If you are at #NeurIPS2024, don’t miss Moshe Eliasof presenting our 𝐆𝐑𝐀𝐍𝐎𝐋𝐀: 𝐀𝐝𝐚𝐩𝐭𝐢𝐯𝐞 𝐍𝐨𝐫𝐦𝐚𝐥𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐆𝐫𝐚𝐩𝐡 𝐍𝐞𝐮𝐫𝐚𝐥 𝐍𝐞𝐭𝐰𝐨𝐫𝐤𝐬….

👇Below we show DiGRAF learns distinct activation functions for different graphs. DiGRAF learns a diffeomorphism using an additional shallow GNN on the input graph and use it as the activation function. We show consistent performance improvement across various tasks and faster

📣Stop by our @NeurIPSConf Poster #3103 Today at 11 AM! for DiGRAF: Diffeomorphic Graph Adaptive Activation Function - an activation function that adapts to the graph structure by learning a diffeomorphism. Joint Work with Xinzhi Wang (equal contrib.), @caromitreka, @brunofmr ,

Graph structure typically influences only node embeddings in GNNs. What if activation functions were graph-adaptive too? Introducing DiGRAF: a novel activation function for GNNs powered by learnable diffeomorphisms. Discover more at our poster #33 tomorrow at @LogConference 🧠📈

This app’s icon is similar to \mathbb{X}.

Entered my buzz cut era. 🙄.

Browsing PASCAL dataset to cheerypick predictions is fun. Lots of meme-worthy images. @CVPR #CVPR2025.