Niv Haim

	Reconstructing Training Data From Real-World Models Trained with Transfer Learning Yakir Oz, Gilad Yehudai, Gal Vardi, Itai Antebi, Michal Irani, Niv Haim SaTML 2026 BibTeX / ArXiv / Code Reconstructing training images from models trained on encodings of popular image backbones (e.g., DINO, CLIP)
	Deconstructing Data Reconstruction: Multiclass, Weight Decay and General Losses Gon Buzaglo, Niv Haim, Gilad Yehudai, Gal Vardi, Yakir Oz, Yaniv Nikankin, Michal Irani NeurIPS 2023 BibTeX / ArXiv / Code / Video Training set reconstruction from multiclass classifiers and models trained with regression loss with some inriguing observations on the implications of weight decay on reconstructability. (Earlier version appeared in ICLR Workshop on Trustworthy ML, 2023)
	SinFusion: Training Diffusion Models on a Single Image or Video Yaniv Nikankin, Niv Haim, Michal Irani ICML 2023 BibTeX / ArXiv / Code / Project Page Diffusion models can be trained on a single image or video, giving rise to diverse video generation and extrapolation.
	Reconstructing Training Data from Trained Neural Networks Niv Haim, Gal Vardi, Gilad Yehudai*, Ohad Shamir, Michal Irani NeurIPS 2022 ORAL BibTeX / ArXiv / Code / Project Page / Video We show that a large portion of the training data can be reconstructed from the parameters of trained MLP binary classifiers. Our method stems from theoretical results about the implicit bias of neural networks trained with gradient descent
	Diverse Generation from a Single Video Made Possible Niv Haim, Ben Finestein, Niv Granot, Assaf Shocher, Shai Bagon, Tali Dekel, Michal Irani ECCV 2022 BibTeX / ArXiv / Code / Video / Project Page We generate diverse video samples from a single video using patch-based methods. Our results outperform single-video GANs in visual quality and are orders of magnitude faster to generate (Extended abstract appeared at AI For Content Creation Workshop @ CVPR, 2022)
	From Discrete to Continuous Convolution Layers Assaf Shocher, Ben Finestein, Niv Haim, Michal Irani Preprint*, 2020 BibTeX / ArXiv Learning continuous convolution kernels improve translation equivariance and allow test time scales augmentations
	Implicit Geometric Regularization for Learning Shapes Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, Yaron Lipman ICML 2020 BibTeX / ArXiv / Code / Video Using an "Eikonal regularization" term with implicit neural representation works surprisingly well for modelling complex surfaces
	Controlling Neural Level Sets Matan Atzmon, Niv Haim, Lior Yariv, Ofer Israelov, Haggai Maron, Yaron Lipman NeurIPS 2019 BibTeX / ArXiv / Code / Poster Making input points differentiable (w.r.t model parameters), and using it for shape modelling, improved robustness to adversrial examples and more
	Surface Networks via General Covers Niv Haim, Nimrod Segol, Heli Ben-Hamu, Haggai Maron, Yaron Lipman ICCV 2019 BibTeX / ArXiv / Code Transforming 3D shapes to image representation so we can feed them to off-the-shelf CNNs and do classification, human-parts segmentation and more
	Extreme close approaches in hierarchical triple systems with comparable masses Niv Haim, Boaz Katz MNRAS 2018 BibTeX / ArXiv / Code Ever wondered if your hierarchical three-body system will eventually collide? find out by plugging your initial conditions into our analytical prediction formula (that works with high probability)

I play the violin [YouTube]

I sometimes write about my travels [blog]

	Reconstructing Training Data From Real-World Models Trained with Transfer Learning Yakir Oz, Gilad Yehudai, Gal Vardi, Itai Antebi, Michal Irani, Niv Haim SaTML 2026 BibTeX / ArXiv / Code Reconstructing training images from models trained on encodings of popular image backbones (e.g., DINO, CLIP)
	Deconstructing Data Reconstruction: Multiclass, Weight Decay and General Losses Gon Buzaglo, Niv Haim, Gilad Yehudai, Gal Vardi, Yakir Oz, Yaniv Nikankin, Michal Irani NeurIPS 2023 BibTeX / ArXiv / Code / Video Training set reconstruction from multiclass classifiers and models trained with regression loss with some inriguing observations on the implications of weight decay on reconstructability. (Earlier version appeared in ICLR Workshop on Trustworthy ML, 2023)
	SinFusion: Training Diffusion Models on a Single Image or Video Yaniv Nikankin, Niv Haim, Michal Irani ICML 2023 BibTeX / ArXiv / Code / Project Page Diffusion models can be trained on a single image or video, giving rise to diverse video generation and extrapolation.
	Reconstructing Training Data from Trained Neural Networks Niv Haim, Gal Vardi, Gilad Yehudai*, Ohad Shamir, Michal Irani NeurIPS 2022 ORAL BibTeX / ArXiv / Code / Project Page / Video We show that a large portion of the training data can be reconstructed from the parameters of trained MLP binary classifiers. Our method stems from theoretical results about the implicit bias of neural networks trained with gradient descent
	Diverse Generation from a Single Video Made Possible Niv Haim, Ben Finestein, Niv Granot, Assaf Shocher, Shai Bagon, Tali Dekel, Michal Irani ECCV 2022 BibTeX / ArXiv / Code / Video / Project Page We generate diverse video samples from a single video using patch-based methods. Our results outperform single-video GANs in visual quality and are orders of magnitude faster to generate (Extended abstract appeared at AI For Content Creation Workshop @ CVPR, 2022)
	From Discrete to Continuous Convolution Layers Assaf Shocher, Ben Finestein, Niv Haim, Michal Irani Preprint*, 2020 BibTeX / ArXiv Learning continuous convolution kernels improve translation equivariance and allow test time scales augmentations
	Implicit Geometric Regularization for Learning Shapes Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, Yaron Lipman ICML 2020 BibTeX / ArXiv / Code / Video Using an "Eikonal regularization" term with implicit neural representation works surprisingly well for modelling complex surfaces
	Controlling Neural Level Sets Matan Atzmon, Niv Haim, Lior Yariv, Ofer Israelov, Haggai Maron, Yaron Lipman NeurIPS 2019 BibTeX / ArXiv / Code / Poster Making input points differentiable (w.r.t model parameters), and using it for shape modelling, improved robustness to adversrial examples and more
	Surface Networks via General Covers Niv Haim, Nimrod Segol, Heli Ben-Hamu, Haggai Maron, Yaron Lipman ICCV 2019 BibTeX / ArXiv / Code Transforming 3D shapes to image representation so we can feed them to off-the-shelf CNNs and do classification, human-parts segmentation and more
	Extreme close approaches in hierarchical triple systems with comparable masses Niv Haim, Boaz Katz MNRAS 2018 BibTeX / ArXiv / Code Ever wondered if your hierarchical three-body system will eventually collide? find out by plugging your initial conditions into our analytical prediction formula (that works with high probability)