Will present this at #CVPR ✈️ See you in Nashville 🇺🇸! Kudos to the team 👏 Antonio A. Gargiulo, @mariasofiab.bsky.social, @sscardapane.bsky.social, Fabrizio Silvestri, Emanuele Rodolà.

Donato Crisostomi ✈️ NeurIPS@crisostomi.bsky.social · 2y ago

📢Prepend “Singular” to “Task Vectors” and get +15% average accuracy for free! 1. Perform a low-rank approximation of layer-wise task vectors. 2. Minimize task interference by orthogonalizing inter-task singular vectors. 🧵(1/6)

📢Prepend “Singular” to “Task Vectors” and get +15% average accuracy for free! 1. Perform a low-rank approximation of layer-wise task vectors. 2. Minimize task interference by orthogonalizing inter-task singular vectors. 🧵(1/6)

📣 Come check it this Friday at #NeurIPS!

Donato Crisostomi ✈️ NeurIPS@crisostomi.bsky.social · 2y ago

I know you're probably thinking, "Yeah, these neuron-permutation-based model merging methods are cool.. but are they cycle-consistent (CC)?" Say no more! It just so happens that our new #NeurIPS24 paper covers exactly this! Huh? No idea what I am talking about? Read on (1/6)

I know you're probably thinking, "Yeah, these neuron-permutation-based model merging methods are cool.. but are they cycle-consistent (CC)?" Say no more! It just so happens that our new #NeurIPS24 paper covers exactly this! Huh? No idea what I am talking about? Read on (1/6)

First blue post (still have to figure out how tweets are called here) 💡idea: we consider task vectors at the layer level and reduce task interference by decorrelating the task-specific singular vectors of any matrix-structured layer 🔬results: large-margin improvements across all vision benchmarks

Simone Scardapane@sscardapane.bsky.social · 2y ago

*Task Singular Vectors: Reducing Task Interference in Model Merging* We show that task vectors are inherently low-rank, and we propose a merging method that significantly improves SOTA. arxiv.org/abs/2412.00081