---

Software engineer specialized in machine learning (especially deep reinforcement learning and planning).

Also have significant experience in embedded systems, high performance computing, hardware/software interfacing, graphics, computer vision, and electronics design.

High proficiency in modern C++.

Professional Experiences

• Engineering for machine learning research

• Interaction Networks for Learning about Objects, Relations and Physics

  • papers.nips.cc/paper/6418-interaction-networks-for-learning-about-objects-relations-and-physics

  • Implemented and ran experiments for simulating physics

• AlphaGo Project

  • deepmind.com/research/alphago

  • Designed and ran experiments in search

  • Worked on scaling the search algorithm to thousands of machines with hundreds of cores each

  • Solved scalability issues with lock-free data structures, atomic access, and partitioning to minimise critical sections

• AlphaZero Project

  • arxiv.org/abs/1712.01815

  • Responsible for the chess version of AlphaZero

  • Devised many experiments and designed many tools to measure and improve performance on chess

• Embedded C++ development with a talented team that puts a lot of emphasis on doing things right (low technical debt)

• Developed firmware for HD IP surveillance cameras with ARM Cortex-A9 CPU on a specialized SoC

• Wrote, maintained, and debugged highly parallel code with complex synchronization requirements and a Linux kernel driver for controlling lenses with electronic focus

• Post-silicon bringup of Kepler generation graphics processor units (GPU)

• Binning and characterization of GPUs (performance spread at different voltages and temperatures)

• Performed high frequency signal integrity, power sequencing, and PLL tests

• Analyzed and "hacked" PCB designs to test power rail sequencing tolerances

• Designed and implemented a self-adjusting vehicle head restraint that senses the occupant's head position and adjust the head restraint to maintain optimal distance in real time to minimize whiplash in case of collision

• Completed the entire electrical system from system design, component spec-ing, schematics/PCB design, assembly, firmware programming (thousands of lines of code including a custom multi-tasking system), to testing, on a very tight schedule

• Designed multiple 2 layer PCBs with analog electromagnetic sensing circuitry, ARM Cortex-M3 (STM32) MCU, brushless motors commutation and control, and USB interfacing

• Delivered a reliable system that successfully met all requirements and provided thorough formal documentation

• Implemented a major feature in a cross-platform C++ game, "Dead Rising 2: Case West", that involved scene analysis, heuristics, front-end, audio, rendering, and animation support

• Collaborated with game designers closely to deliver the feature exactly as envisioned, on a very tight schedule, with constantly evolving requirements

• Developed C# tools to improve QA, artists, and programmers efficiency

Technical Projects

• Used multiple deep neural networks with temporal-difference reinforcement learning to play chess

• Implemented neural networks and stochastic gradient descent (AdaDelta) in C++11 using the Eigen BLAS library

• Engineered a smooth low level feature representation that resulted in highly efficient learning with good generalization

• Devised a novel way to combine the power of minimax searches in stochastic situations with the power of neural networks in pattern recognition

• The resulting program plays at the level of FIDE International Master after 72 hours of training - by far the highest strength ever achieved using machine learning from low level features

• Results were widely reported in both mainstream and tech media, including the MIT Technology Review

• Thesis: http://arxiv.org/abs/1509.01549

• Used deep learning techniques to segment 3D brain MRI images for Alzheimer's diagnosis

• Implemented 3 convolutional architectures in Theano (Python meta-programming framework) with different computational power vs accuracy tradeoffs, and performed experiments on GPUs

• Wrote a report summarizing the development history of neural networks, including recent discoveries such as ReLU activation, DropOut, model compression ("Dark Knowledge"), and convolutional architectures

• Report: http://arxiv.org/abs/1505.02000

Education

Interests

• Making things - woodworking, electronics design, art installations

• Backpacking, camping, social psychology and mental health, musical theatre, judo, fish-keeping, chess, SCUBA diving

• Certified private pilot (airplane) in the US and Canada