Jonathan Svirsky

About

My research focuses on [multimodal] representation learning for efficient and interpretable neural networks. I design lightweight speech models, adaptive feature selection methods, and efficient training techniques for LLMs. I’m passionate about bridging theory and real-world impact, developing deployable and efficient AI systems grounded in data-driven modeling.

Research Interests

• Efficient Speech Processing and Generation
• Multimodal Representation Learning
• Adaptive Feature Selection and Dynamic Sparsification

Publications

• COPER: Correlation-based Permutations for Multi-View Clustering
  R. Eisenberg, J. Svirsky, O. Lindenbaum
  ICLR, 2025
• AdaRankGrad: Adaptive Gradient-Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning
  Y. Refael, J. Svirsky, B. Shustin, W. Huleihel, O. Lindenbaum
  ICLR, 2025
• Sparse Binarization for Fast Keyword Spotting
  J. Svirsky, U. Shaham, O. Lindenbaum
  Interspeech, 2024
• Interpretable Deep Clustering for Tabular Data
  J. Svirsky, O. Lindenbaum
  ICML, 2024
• SG-VAD: Stochastic Gates Based Speech Activity Detection
  J. Svirsky, O. Lindenbaum
  ICASSP, 2023
• Deep Unsupervised Feature Selection by Discarding Nuisance and Correlated Features
  U. Shaham, O. Lindenbaum, J. Svirsky, Y. Kluger
  Neural Networks 152, 34–43, 2022
• Discovery of Single Independent Latent Variable
  U. Shaham, J. Svirsky, O. Katz, R. Talmon
  NeurIPS 36, 2022
• Differentiable Unsupervised Feature Selection Based on a Gated Laplacian
  O. Lindenbaum, U. Shaham, E. Peterfreund, J. Svirsky, N. Casey, Y. Kluger
  NeurIPS 34, 1530–1542, 2021

Projects & Code

• ICV - Provable Speech Attributes Conversion via Latent Independence
• Interpretable Deep Clustering - Interpretable Deep Clustering for Tabular Data
• SparkNet - Sparse Binarization for Fast Keyword Spotting
• SG-VAD - Stochastic Gates Based Speech Activity Detection
• Audio Samples - Discovery of Single Independent Latent Variable

Visitors Map