I am a research scientist at Google AI Perception in Mountain View, California. I am always learning, and also trying to help machines learn, faster and better. My main research interests vary over time and include deep-network structure learning, probabilistic modeling, robust multi-label classification and weakly supervised learning.
In the past I was more interested in Quantum Computation, a field in which I completed my M.Sc. under the supervision of Prof. Dorit Aharonov, graduating magna cum laude. Before that, I completed my B.Sc. degree in Computer Science and Computational Biology summa cum laude.
Neural Network Structure Learning
The design of deep neural networks has traditionally been a tedious, manual process of trial and error strongly powered by intuition and lacking rigorous methodologies. On the other hand, recent powerful automated architecture search algorithms tend to be prohibitively expensive to train, and their results depend on a manually well crafted search space and proxy tasks which are, again, based on intuition.
To address this we invented MorphNet, a simple Architecture Search algorithm which learns the number of neurons in each layer jointly with the model weights - in a single training run. The approach has minimal overhead which allows it to be applied to Billion-item datasets and is flexible enough to handle (almost) any type of deep architecture. The method can be set to target specific constrained resources such as FLOPs or model size when designing the architecture.
The MorphNet algorithm was used to speed-up and improve architectures in various modalities on academic benchmarks, and is used by multiple projects within Google.
MorphNet: Fast & Simple Resource-Constrained Structure Learning of Deep Networks. Ariel Gordon, Elad Eban, Ofir Nachum, Bo Chen, Hao Wu, Tien-Ju Yang, Edward Choi. CVPR 2018.
Code [NEW March 2019]
Get What You Train For: Learning With Objectives That Match Your Evaluation Metrics
Almost all machine learning classification models are optimized for classification accuracy, via a cross entropy surrogate. However, most often the real objective one cares about is different and can be precision at a fixed recall, recall at a fixed precision, precision-recall AUC, ROC AUC or similar metrics. These metrics are non-decomposable (because they depend on how the model classifies the dataset as a whole and do not decouple across data-points), and have no derivatives so they cannot be directly used in training.
Towards bridging the gap between the training objective and evaluation metric we derived simple, differentiable surrogates for these metrics that rely on convex relaxations and a bit of optimization.
Models trained with our loss functions are used by dozen of teams across Google, improving image classification, spam and phishing detection and other products.
Scalable Learning of Non-Decomposable Objectives. Elad Eban, Mariano Schain, Alan Mackey, Ariel Gordon, Ryan Rifkin, Gal Elidan. AISTATS 2017.
In a follow up work we used quantile estimation and provided simple and effective loss functions for precision at k and related metrics.
Constrained Classification and Ranking via Quantiles. Alan Mackey, Xiyang Luo, Elad Eban. arxiv 2018
A cortical–hippocampal–cortical loop of information processing during memory consolidation. Gideon Rothschild, Elad Eban, Loren M Frank. Nature Neuroscience 2017.
Interactive Proofs for Quantum Computations
Interactive Proofs For Quantum Computations. Dorit Aharonov, Michael Ben-Or, Elad Eban. ICS 2008.