Ofir Lindenbaum @ Bar-Ilan University

Publications

Group highlights

For a full list see below or go to Google Scholar.

GroupFS method figure from the paper

Unsupervised Feature Selection Through Group Discovery

GroupFS jointly discovers feature groups and selects the informative ones without labels, turning feature selection into a structured, graph-smooth learning problem.

S. Lifshitz, O. Lindenbaum, G. Mishne, R. Meir, H. Benisty

AAAI 2026 Oral

Entropy-adaptive permutation learning figure from the paper

Learning Permutation from Structure Without Supervision

An entropy-adaptive Gumbel-Sinkhorn formulation learns hidden orderings from structural objectives, sharpening confident assignments while leaving ambiguous ones flexible.

R. Eisenberg, O. Lindenbaum

ICML 2026

Structured sparsity fine-tuning method figure from the paper

Train Less, Infer Faster: Efficient Model Finetuning and Compression via Structured Sparsity

The method learns structured row and column gates that sparsify foundation models during adaptation, reducing inference cost while keeping the base weights frozen.

J. Svirsky, Y. Refael, O. Lindenbaum

AISTATS 2026

Partition First, Embed Later method figure from the paper

Partition First, Embed Later: Locally Linear Manifold Embedding of Partitioned High-Dimensional Data

A feature-partitioning step separates mixed signals before embedding, revealing local low-dimensional structure that can be blurred by a single global projection.

E. Peterfreund, J. Gradstein, O. Lindenbaum, Y. Kluger, R. Landa

ICML 2025 Oral

AdaRankGrad memory and rank figure from the paper

AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning

AdaRankGrad tracks the effective rank of gradient updates and allocates low-rank optimizer memory where it is useful, making LLM training and fine-tuning lighter.

Y. Refael, J. Svirsky, B. Shustin, W. Huleihel, O. Lindenbaum

ICLR 2025

SUMO moment conditioning figure from the paper

SUMO: Subspace-Aware Moment-Orthogonalization for Accelerating Memory-Efficient LLM Training

SUMO orthogonalizes optimizer moments inside a dynamically adapted low-dimensional subspace, accelerating memory-efficient LLM training beyond low-rank savings alone.

Y. Refael, G. Smorodinsky, T. Tirer, O. Lindenbaum

NeurIPS 2025

COPER method figure from the paper

COPER: Correlation-Based Permutations for Multi-View Clustering

COPER aligns views by learning within-cluster permutations from correlation structure, producing stronger shared representations for multi-view clustering.

R. Eisenberg, J. Svirsky, O. Lindenbaum

ICLR 2025 Spotlight

Stochastic Gates method figure from the paper

Feature Selection Using Stochastic Gates

Stochastic Gates make feature selection differentiable: each variable receives a learnable noisy gate, so compact and interpretable subsets emerge during training.

Y. Yamada, O. Lindenbaum, S. Negahban, Y. Kluger

ICML 2020

Locally Sparse Neural Networks method figure from the paper

Locally Sparse Neural Networks for Tabular Biomedical Data

Local sparse subnetworks choose different variables for different samples, preserving predictive accuracy while giving sample-specific explanations in tabular biomedical data.

J. Yang, O. Lindenbaum, Y. Kluger

ICML 2022 Spotlight

Geometry-Based Data Generation result figure from the paper

Geometry-Based Data Generation

A graph-geometry view of augmentation generates samples that preserve manifold neighborhoods, improving downstream recovery of latent biological structure.

O. Lindenbaum, J. S. Stanley III, G. Wolf, S. Krishnaswamy

NeurIPS 2018 Spotlight

Full List of publications

Complete bibliography merged from Google Scholar and DBLP. Published versions replace matching preprints when available.

Loading publications