Yoni Choukroun

I am a team lead AI research scientist at Huawei Israel Research Center. I completed my Ph.D. in Computer Science at Tel Aviv University, working in the Deep Learning Lab under the supervision of Prof. Lior Wolf. I obtained my BSc in Computer Engineering from the Technion and my M.Sc. in Computer Science from the Technion under the supervision of Prof. Ron Kimmel.

My research interests are centered around machine/deep learning, computer vision, information theory, and numerical optimization.

Email  /  Google Scholar  /  Github  /  Linkedin

We propose a new method for the weighted aggregation of the gradients in the syncrhonous distributed data parallel setting

We propose the acceleration of the ECCT via adapted ternary quantization and code aware multi-heads to reach Belief Propagation complexity

We propose a new structure learning of the underlying Bayesian graph via a learned masking of the factors' connectivity

We learn new binary linear block codes by cotraining the code together with a new SOTA Transformer neural decoder.

We propose a new foundational decoder that is code, length, and rate invariant and able to reach SOTA even on zero-shot scenarios.

We outperform existing solutions by overcoming several quantum ECC challenges with a deep-learning-based decoder.

We learn the HRF time series via a geometry-aware transformer applied to the neurovascular modalities.

We modelize the channel corruption and the decoding process as a forward and reverse diffusion process.

We employ an adapted Transformer in order to incorporate information about the code and to allow efficient cross-analysis of the channel's output elements.

We develop a novel Transformer based architecture able to learn atomic interactions and we propose a novel molecular data augmentation strategy.

We propose a meta-learning based approach for the selection of the directions defining the suspace span.

We present a proximal subspace optimization algorithm coupled with a second-order oracle backtracking line-search for the optimization of saddle-point problems.

We develop post-training quantization of neural networks to up to 4 bits using an optimal MSE minimization, dual kernel refinement and differentiable finetuning of the quantizatio parameters.

We propose the analysis and the adaptation of the Hamiltonian operator to the field of shape analysis.

We propose the compression of 3D surfaces using sparse matching pursuit over the mesh dictionary built upon the Hamiltonian Operator eigenspace.

We present a preprocessing method of the channel output in order to make the model robust to codeword overfitting.

We develop a novel unsupervised clustering method by proposing a novel discriminative encoding objective and a similarity aware clustering phase.

We propose the compression of 3D meshes using sparse matching pursuit over the mesh dictionary.

We develop a patch-based multiple instance learning method for mammogram classification.

We propose a novel IRLS approximation of the L1 on manifolds equivalent to a constrained iterative Hamiltonian eigendecomposition.

This page design is based on a template by Jon Barron.