Ofer Dekel

Ofer Dekel is a machine learning researcher whose work spans online learning theory, efficient neural-network inference and embedded AI. He earned a computer science PhD at the Hebrew University of Jerusalem under Yoram Singer and coauthored "Online Passive-Aggressive Algorithms" in 2006. In 2015, with Sébastien Bubeck, Tomer Koren and Yuval Peres, he proved that adversarial bandit convex optimization in one dimension has minimax regret Θ̃(√T), and in 2017 he led a Redmond and Bangalore team building the Embedded Learning Library for bread-crumb size processors. He coauthored the Maia 200 AI accelerator paper in August 2026. As of October 2026, he is a Distinguished Engineer at Microsoft.

Now at: Microsoft Research
Born: 1975
Works: 66
Citations: 5,226
h-index: 30
Badges: Bridge

Years link to their sources. Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2006

Online passive-aggressive algorithms

At the Hebrew University of Jerusalem, where he earns a computer science PhD under Yoram Singer, he coauthors "Online Passive-Aggressive Algorithms" in 2006. The Journal of Machine Learning Research paper, written with Koby Crammer, Joseph Keshet, Shai Shalev-Shwartz and Yoram Singer, presents a family of margin-based online learning algorithms for categorization, regression and sequence prediction.
2010–2012

Distributed and adaptive online learning

In 2010, he coauthors "Optimal Distributed Online Prediction using Mini-Batches" with Ran Gilad-Bachrach, Ohad Shamir and Lin Xiao, which converts serial gradient-based online prediction algorithms into distributed ones. In 2012, with Raman Arora and Ambuj Tewari, he defines policy regret for online bandit learning against adaptive adversaries at the International Conference on Machine Learning.
2013

UW/MSR Machine Learning Day

In 2013, he co-organizes the UW/MSR Machine Learning Day with Emily Fox and Ben Taskar, and nearly 300 researchers from Microsoft and the University of Washington take part. At Microsoft Research, he works on applying machine learning to the Bing search engine.
2015

Bandit convex optimization in one dimension

With Sébastien Bubeck, Tomer Koren and Yuval Peres, he proves in 2015 that the minimax regret of adversarial bandit convex optimization in one dimension is Θ̃(√T), which partially resolves a decade-old open problem.
2017

The Embedded Learning Library

In 2017, he leads a multidisciplinary team of computer scientists, software engineers and interns in Redmond and Bangalore, India, that builds machine-learning software and tools to embed AI onto bread-crumb size processors. An early preview of the team's Embedded Learning Library (ELL) goes onto GitHub.
2017

Adaptive networks for efficient inference

With Tolga Bolukbasi, Joseph Wang and Venkatesh Saligrama, he coauthors "Adaptive Neural Networks for Efficient Inference" in 2017. The paper's early-exit and network-selection schemes reach up to a 2.8x speedup on ImageNet networks with under 1% loss of top-5 accuracy.
August 2026

The Maia 200 accelerator paper

In August 2026, he is among the 17 authors of "Maia 200: A Software Defined Dataflow System for Large-scale AI Acceleration". The paper introduces an AI accelerator that delivers 5072 Tflop/s FP8 within a 750W TDP and 7 TB/s of HBM bandwidth.
As of October 2026

Distinguished Engineer at Microsoft

He is a Distinguished Engineer at Microsoft, based in Redmond, Washington.