Michael Stanley

I am a senior machine learning engineer at Zoox where I design, build, train, and deploy learned models for autonomous driving. I previously held a similar role at Parallel Domain where we generated and studied the benefits of synthetic data for computer vision models.

I am also a machine learning advisor to Portrait Creative Network. Here is a demo of Portrait's AI search tool for finding creative talent.

I completed my Masters at the Center for Data Science at New York University. While at NYU, I was a researcher and team leader in the Urban Modeling Group under Prof. Debra Laefer and investigated inverse approaches to 3D shape extraction under Prof. Carlos Fernandez-Granda. Here is my CV.

Before NYU, I spent twelve years in industry, specifically in management consulting at Bain & Company, private equity at CIVC Partners, and product management in IoT security at Symantec Corporation and machine learning and data software at Enigma Technologies.

I studied Mechanical Engineering and Economics at Duke University. I am originally from Lexington, Kentucky.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

Papers

RealDriveSim: A Realistic Multi-Modal Multi-Task Synthetic Dataset for Autonomous Driving

This paper introduces a realistic multi-modal synthetic dataset for autonomous driving that supports both 2D computer vision and LiDAR-based perception tasks with annotations across up to 64 classes. We demonstrate the dataset’s effectiveness through comprehensive evaluation across multiple applications and domains, reporting performance improvements compared to existing synthetic benchmarks.

Metrics for Aerial, Urban LiDAR Point Clouds

This paper introduces five new density and accuracy metrics for aerial point clouds that address the complexity and objectives of modern, dense laser scans of urban scenes.

Bandit Modeling of Map Selection in Counter-Strike: Global Offensive

This paper quantifies and improves upon inefficiencies in the map selection process in e-sports, specifically Counter-Strike Global Offensive.

Blog Posts

Research at Parallel Domain is typically shared via blog posts rather than papers. I led the following experiments.

Boosting Optical Flow with PD Synthetic Data

Parallel Domain (PD) synthetic data improves performance on optical flow tasks by 18.5% by matching flow magnitudes.

Not all Synthetic Datasets are Created Equal

Parallel Domain data improves unsupervised domain adaptation performance by 30% vs. GTA with no changes to model architecture.

Miscellaneous Projects

A patent and some cocktail data science.

quantiacs-ml: An ML Framework for Algorithmic Trading

I started playing with Quantiacs but realized their ML training and evaluation infrastructure is bad. This is a basic framework for training, evaluating, and implementing ML-based trading strategies on the Quantiacs platform.

The Minimalist Cocktail Map

My wife and I created a map of 30 cocktails you can make with just 10 bottles. The drinks are laid out based on the first two principal components over 30+ drink features. Design by Susan Taylor Studios.

US Patent 10,348,758

Patent title: Systems and Methods for Providing Interfaces For Visualizing Threats Within Networked Control Systems. A user interface design patent from the product Anomaly Detection for Industrial Control Systems, launched while at Symantec. The concept was displaying alerts in various subnets within a highly segmented, hierarchical ICS system.

Design and source code from Leonid Keselman's Jekyll implementation of Jon Barron's website