gal vardi

[Nirmit Joshi @nirmitj_]

Work with excellent collaborators, especially Roey Magen. And also @galvardi @nikostsilivis and Nati Srebro. Link: https://t.co/GmuQtHJccA

[Nirmit Joshi @nirmitj_]

Imagine learning from multiple thinkers with different thought processes. We find that passive data from just two thinkers can be cryptographically hard to learn from. But active collection allows arbitrarily high accuracy, with per-thinker CoT independent of target accuracy.

[Nirmit Joshi @nirmitj_]

Happy to share our work on:
"A Theory of Learning with Autoregressive Chain of Thought"
https://t.co/40npwsrNRv

Joint work with great collaborators: @galvardi, Adam Block, @SurbhiGoel_, @zhiyuanli_
, Theodor Misiakiewicz, Nati Srebro

[Julia Kempe @KempeLab]

Optimization induces implicit bias. We study general steepest descent in homogeneous nets & show (generalized) convergence to a (generalized) KKT pt. Adam presents a curious case between l2 & l1: https://t.co/uUev7gTLCG
With Nikos Tsilivis & @galvardi @NYUDataScience @AIatMeta

[Spencer Frei @sfrei_]

New preprint with @galvardi: https://t.co/AKguvWEzKB
We prove task- & sample-complexity bounds for in-context learning in transformers for classification tasks, focusing on a 1-layer linear attention architecture. We find they can exhibit benign overfitting in-context!

[Naftali Bennett נפתלי בנט @naftalibennett]

An entire family murdered in cold blood.

Kedem family: Father Jonathan, mother Tamar, 6-year-old girls Shachar and Arbel, and 4-year-old boy Omer.

Look at their happy faces.
Their love.

All of them murdered by Palestinian terrorists at Nir-oz kibbutz.
Just because they’re Jews.
🇮🇱💔

[Niv Haim @HaimNiv]

Glad to share our new work: "Deconstructing Data Reconstruction: Multiclass, Weight Decay and General Losses" https://t.co/KEHM6bSJy5

About trainset reconstruction in multiclass & regression tasks + perks

(w/ @GonBuzaglo @Giladude @galvardi @ozyakir @YNikankin M. Irani)
🧵1/5

[Niv Haim @HaimNiv]

Thrilled to share that our paper "𝘙𝘦𝘤𝘰𝘯𝘴𝘵𝘳𝘶𝘤𝘵𝘪𝘯𝘨 𝘛𝘳𝘢𝘪𝘯𝘪𝘯𝘨 𝘋𝘢𝘵𝘢 𝘧𝘳𝘰𝘮 𝘛𝘳𝘢𝘪𝘯𝘦𝘥 𝘕𝘦𝘶𝘳𝘢𝘭 𝘕𝘦𝘵𝘸𝘰𝘳𝘬𝘴" was accepted to #NeurIPS 2022 as an 𝗼𝗿𝗮𝗹 𝗽𝗿𝗲𝘀𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻!
w/ @Giladude @galvardi

Webpage: https://t.co/AONNAPY7dG
🧵1/n

[Spencer Frei @sfrei_]

New preprint with @galvardi, Peter Bartlett, Nati Srebro, and @weihu_ on the implicit bias of gradient descent/flow (GD/GF) in two-layer leaky ReLU networks when trained on high-dimensional data: https://t.co/cNYKlDsIBc

[Gilad Yehudai @Giladude]

Check out our latest work where we reconstruct large portions of the actual training data from trained neural networks. Joint work with @galvardi @HaimNiv

Project page: https://t.co/HN4q3m5vGz
abs: https://t.co/bPc5LP00fC