Christopher Serrano

ML Researcher / AI Engineer / Mad Scientist
GPU-Accelerated Deep RL | Sim2Real | Trusted Autonomy

End-to-end Reinforcement Learning for real-world autonomous platforms.

Featured Work

L-29 jet on tarmac The War Zone

AI-Enabled L-29 Jets Fly Mock Air-To-Air Mission
ECHOS flight test Lockheed Martin

Airborne Battle Management of AI-Controlled Aircraft
Two L-29 jets performing simulated intercept Lockheed Martin

AI for Air-to-Air Tactical Intercepts

DARPA Assured Autonomy — Autonomous MRZR with formally verified perception and Deep RL control

Experience

Lockheed Martin Skunk Works

May 2023 – Present

Aeronautics, Advanced Development Programs

Senior Staff AI / Machine Learning Engineer L5

HRL Laboratories

Aug 2019 – May 2023

Intelligent Systems Laboratory, Operational Autonomy Center

Scientist V (July 2021 – May 2023)

  • Led and supported multiple CRAD proposals, including as co-PI on a DARPA BAA
  • Developed custom GPU-accelerated RL training environments that scaled simulation 10,000x
  • Developed novel efficient Transformer architectures for processing formal logics
  • Led and supported multiple IRAD proposals annually with a 67% win rate

Scientist IV (July 2020 – June 2021)

  • Developed real-time Deep RL adversarial AI attack algorithm (RTA3) with 91% attack success rate
  • Integrated formally verified neural network perception and Deep RL controllers into autonomous MRZR for DARPA Assured Autonomy
  • Developed augmented reality adversarial AI attack against 9DoF autonomous vision systems

Post Masters Scientist (Aug 2019 – June 2020)

  • Developed formal verification methodology (Generate & Verify) for neural network perception systems
  • Built neural network LiDAR perception systems with formally verified performance guarantees

Graduate Research Intern (May 2018 – July 2019)

  • Developed Introspection Learning algorithm — 99.6% reduction in failures during training
  • Built custom RL training environments in Python and C++ with PyBullet, Unreal Engine, and AirSim

6 patent applications · 8 invention disclosures · 6 paper submissions

Georgia Institute of Technology

2017 – Present

CS 7642 Reinforcement Learning and Decision Making

Instructional Associate (May 2019 – Present)

  • Automated plagiarism detection utilizing MOSS document similarity

Teaching Assistant (March 2017 – April 2019)

Publications

Featured Point cloud sphere visualization

Assured Point Cloud Perception

Serrano, C.R., Nogin, A., Warren, M.A. (2023)

Demonstration of verification of global correctness for neural network point cloud perception models performing regression and classification.

Presented at the 2023 IEEE International Conference on Assured Autonomy

View publication

Introspection Learning

Serrano, C.R., Warren, M.A. (2019)

Synthesizes experience without requiring environment interaction by asking the policy directly about situations and actions, incorporating formal verification artifacts into Deep RL training.

Presented at the 2019 AAAI Spring Symposium on Verification of Neural Networks

View on arXiv

RTA3: A Real Time Adversarial Attack on Recurrent Neural Networks

Serrano, C.R., Sylla, P., Gao, S., Warren, M.A. (2020)

General application of deep reinforcement learning to the generation of periodic adversarial perturbations in a black-box approach to attack recurrent neural networks.

Presented at the 2020 IEEE Deep Learning and Security Workshop

View publication
Self-Satisfied framework figure

Self-Satisfied: An End-to-End Framework for SAT Generation and Prediction

Serrano, C.R., Gallagher, J., Yamada, K., Kopylov, A., Warren, M.A. (2024)

GPU-accelerated SAT problem generation and a novel Satisfiability Transformer (SaT) architecture with head slicing for reducing sequence length. Demonstrated on problems with thousands of variables from SAT Competition 2022.

arXiv preprint

View on arXiv

Generate and Verify: Semantically Meaningful Formal Analysis of Neural Network Perception Systems

Serrano, C.R., Sylla, P., Warren, M.A. (2020)

A notion of global correctness for neural network perception models performing regression with respect to a generative neural network with a semantically meaningful latent space.

arXiv preprint

View on arXiv

Patents

Granted

Solving based introspection to augment the training of reinforcement learning agents for control and planning on robots and autonomous vehicles

Warren, M.A., Serrano, C.R. (2020) — US Patent 11,669,731

Granted

Automated system for generating approximate safety conditions for monitoring and verification

Heersink, B.N., Warren, M.A., Serrano, C.R. (2021) — US Patent 11,663,370

Granted

Deep reinforcement learning based method for surreptitiously generating signals to fool a recurrent neural network

Warren, M.A., Serrano, C.R., Sylla, P. (2021) — US Patent 12,073,318

Application

Neural network architecture for small lidar processing networks for slope estimation and ground plane segmentation

Serrano, C.R., Warren, M.A., Nogin, A. (2021) — US Patent App. 16/950,803

Application

Deep reinforcement learning method for generation of environmental features for vulnerability analysis and improved performance of computer vision systems

Warren, M.A., Serrano, C.R. (2021) — US Patent App. 17/115,646

Application

Method for proving or identifying counter-examples in neural network systems that process point cloud data

Warren, M.A., Serrano, C.R., Nogin, A. (2021) — US Patent App. 17/078,079

Education

Georgia Institute of Technology

MS, Computer Science

Specializing in Machine Learning

2019

University of California
at Santa Barbara

BA, Political Science

Minors in History and Art History

2004