Biography
Eugene Brevdo is the Tech Lead/Manager of the Learned Systems group at Google Brain. His research interests span several interconnected areas:
- Software systems for training and deploying ML, Bandits, and RL models.
- Uncertainty, evolutionary strategies, and population-based training.
- Machine Learning applied to optimizing large software systems (databases, datacenter scheduling, caches, compilers like LLVM and XLA/TPU).
Eugene received his PhD in Electrical Engineering from Princeton University, where his advisers were Peter Ramadge and Ingrid Daubechies.
-
PhD in Electrical Engineering, 2011
Princeton University
-
BSc in Electrical, Computer, and Systems Engineering, 2005
Rensselaer Polytechnic Institute
Experience
Co-TLM of the TF-Agents team (2018 - 2021).
TLM of the Brain Learned Systems Team (2017 - now). Clients include Spanner, Compiler, and Cloud infrastructure teams.
- Built smarter query optimizers, cache eviction algorithms, inlining and register allocation passes.
- Grew the Learned Systems team from 1 to 7 researchers and engineers.
- Aligned engagements between Brain, Technical Infrastructure, and Cloud orgs.
- Set research direction for systems and ML engineers.
SWE on Brain Applied Machine Intelligence team.
- Core TensorFlow maintainer.
- Developed interfaces and support for sparse and sequential input, debugged graph control flow, implemented CPU and GPU kernels; whatever needed doing.
- Founding SWE / API designer of TF Distributions (now Tensorflow Probability).
I worked on two teams:
- Computational Climatology: Statistical weather forecasting in the short-to-medium-term scale (2 weeks-2 years) using a combination of techniques from climatology, machine learning/statistics, and spatiotemporal signal processing.
- Computational Agronomy: Analyzed, assimilated, and reconciled remotely sensed weather and agricultural data. Built growth forecasts for corn, sorghum, soy, and winter wheat.
Focused on applications of Compressive Sensing to inverse problems in medical imaging.
- Developed CS-based estimator for Computational Tomography with Sinogram Occlusion.
- Developed a novel CS-based reconstruction technique for Ultrasound tomography.
Projects
MATLAB Synchrosqueezing Toolbox
This toolbox implements several time-frequency and time-scale analysis methods, including the Forward/Inverse Discretized CWT, and CWT-based Synchrosqueezing.
Reverb
Reverb is an efficient and easy-to-use data storage and transport system designed for machine learning research. Reverb is primarily used as an experience replay system for distributed reinforcement learning algorithms but the system also supports multiple data structure representations such as FIFO, LIFO, and priority queues.