Itay Hubara

Itay Hubara is a senior research scientist at Habana Labs, a company that develops AI accelerators for deep learning. Itay received his PhD from the Technion-Israel Institute of Technology. Itay research is focused on deep neural network acceleration for tasks such as generative AI, NLP, speech, and computer vision. He is one of the contributors to the MLPerf inference and training benchmarks, a standard for measuring the performance of machine learning systems. He has published several papers in top conferences and journals, such as NeurIPS, CVPR, ICLR, ICML, and JMLR, and has received more than 10,000 citations for his work. His research interests include deep learning, computer vision, and hardware-software co-design.

CV

Education

B.Sc. in Electrical Engineering, Technion Israel Institude of Technology
M.Sc. in Electrical Engineering (summa cum laude), Technion Israel Institude of Technology
Ph.D iin Electrical Engineering, Technion Israel Institude of Technology
Thesis title: Toward Fast and efficent Deep Learning

Work experience

2022- present Habana Labs and Intel; company, Principal engineer, Senior Researcher
- 2016-2021 Habana Labs, Senior Researcher (US branch) • Research and development- focused on Compressing and speeding up Deep Neural Network algorithms. • Leading Habana’s MLPerf submissions and benchmarking efforts. • Technical lead and mentor for several members of the algorithm and workload analysis teams
2014-2016 Intel LTD, Algorithm Engineer - Computer Vision Team Research and development- focused on Deep Neural Network algorithms (CNN, RNN) and global path planning in high dimensional space algorithms
2014-2016 Technion - Israel Institute of Technology, Teaching Assistant & B.Sc. Final Project Supervisor *Taught Machine learning/image processing courses and labs to undergraduate students and supervised deep learning- oriented B.Sc. projects
2014-2015 AerialGuard, Algorithm Expert – External Advisor Created a widely used fast collision avoidance algorithm with low time complexity
2010-2014 CSR Group LTD, Algorithm Engineer - Computer Vision Team Research and development- focused on DFD (Depth From De-focus), object detection and tracking

Skills

Algorithm
- Machine learning/ Deep learning
- Neural Languge Processing: LLMs starting from older versions sucha as LSTM to the newer decoder basde models (Llama/MoE/GPT)
- Computer vision: Detection,segmentation recognition
- Optimization: Quantization (FP8/FP4), sparsification (block-sparse,unstructured), LLMs inference acceleration serving solutions (vLLM/TGI)
- Combinatorical problems (Integer\Linear) programing
Coding
- Python, Pytorch, TF, Cuda, C++
- Github fouent
MLPerf (training/inference)
- Benchmark owner (old SSD/ Llama-70B-Lora)
- Leading Habana efforts and representing the company in all MLPerf meetings

Awards

2021 - MLIS scholarship for excellent graduate student in data science. 2019 – ICML top 5% reviewers award 2018 – Jury Award for excellent graduate student 2017 – AI Grant 1.0 2017 – MSc Electrical Engineering with highest honor – summa cum laude

Publications

E Kinderman, I Hubara, H Maron, D Soudry Foldable SuperNets: Scalable Merging of Transformers with Different Initializations and Tasks. arxiv preprint 2024

Y Blumenfeld, I Hubara, D Soudry Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators ICLR 2024

• Chmiel, B., Hubara, I., Banner, R., & Soudry, D. (2022). Optimal Fine-Grained N: M sparsity for Activations and Neural Gradients. ICLR 2023.

• Hubara, I., Chmiel, B., Island, M., Banner, R., Naor, S., Soudry, D., Accelerated Sparse Neural Training: A Provable and Efficient Method to Find N: M Transposable Masks. NeurIPS 2021

• Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., & Soudry, D. A_ccurate Post Training Quantization With Small Calibration Sets._ ICML 2021

• Hoffer, E., Ben-Nun, T., Hubara, I., Giladi, N., Hoefler, T., & Soudry, D. Augment Your Batch: Improving Generalization Through Instance Repetition. CVPR 2020

• Haroush, Matan, Itay Hubara, Elad Hoffer, and Daniel Soudry. The knowledge within: Methods for data- free model compression. CVPR 2020.

• Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry. Scalable methods for 8-bit training of neural networks (NIPS 2018)

• Elad Hoffer, Itay Hubara, and Daniel Soudry. Fix your classifier:the marginal value of training the last weight layer (ICLR 2018)

• Hoffer, E., Hubara I., & Soudry, C. D. T_rain longer, generalize better: closing the generalization gap in large batch training of neural networks_ (NIPS 2017)

• Hubara I., Courbariaux, M., Soudry, C. D., El-Yaniv, R., & Bengio, Y. Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations (JMLR 2017)

• Hubara I., Courbariaux, M., Soudry, C. D., El-Yaniv, R., & Bengio, Y. Binarized Neural Networks: Training Neural Networks with Weights and Activations Constrained to + 1 or −1 (NIPS 2016)

• Bhonker, Nadav, Shai Rozenberg, and Itay Hubara. Playing SNES in the Retro Learning Environment. (ICLR 2016 workshop).

• Soudry, D., Hubara, I., & Meir, R. (2014). Expectation backpropagation: parameter-free training of multilayer neural networks with continuous or discrete weights. (NIPS 2014)

• Additional publication can be found at: https://scholar.google.com/citations?user=dyYryZYAAAAJ&hl=en

Patents

• Hubara, Itay. Large-scale computations using an adaptive numerical format. U.S. Patent No. 10,491,239. 26 Nov. 2019. • El-Yaniv, Ran, Hubara, Itay, and Soudry Daniel. Quantized neural network training and inference. U.S. Patent Application No. 15/478,531.