About Me
I am a PhD candidate at the Karlsruhe Institute of Technology working at the intersection of machine learning, signal processing, and experimental physics, with a focus on methods that translate reliably from theory into real systems. Previously, I studied physics at Brandeis University and electrical and computer engineering at Olin College, and worked on CMS-related projects at MIT, Fermilab, and CERN. Through these experiences, I developed a strong appreciation for end-to-end systems: filtering pipelines, real-time reconstruction, and distributed computation that must perform reliably in production-like settings.
Current Research Focus
Signal processing pipelines where ML models are used to replace or complement specific classical stages (e.g. template matching or feature extraction) under fixed latency and noise constraints.
Designing hybrid pipelines that integrate neural sequence models (e.g., attention-based architectures)
with optimal filtering and Fourier-domain methods to reconstruct physical quantities from noisy,
high-rate time-series data.
Anomaly detection on highly imbalanced experimental data, with a focus on controlling false positives and understanding failure modes rather than purely maximizing detection rate.
Developing unsupervised and weakly supervised methods—such as PCA-based models, autoencoders,
and density estimation techniques—to identify rare, nonphysical, or pathological signal morphologies
in large datasets.
ML models trained with explicit physical constraints or priors, where we studied the trade-offs between model flexibility, stability during training, and bias introduced by imperfect physical assumptions.
Combining domain knowledge from signal processing and statistical physics with learnable components
to improve robustness, interpretability, and sample efficiency relative to purely black-box approaches.
Supporting infrastructure for scientific ML workflows, including data handling, experiment reproducibility, and resource scheduling, primarily in research-scale environments rather than production systems.
Building containerized and distributed workflows using tools such as HTCondor, JupyterHub, and GPU clusters
to support reproducible training, large-scale inference, and rapid iteration in collaborative environments.
Technical Interests
My interests lie at the boundary between physical systems and computation, particularly where models must operate under real-world constraints.
- Representation learning for time series, including self-supervised learning on large volumes of unlabeled data
- Simulation-based and likelihood-free inference, with emphasis on flow-based and amortized methods
- Signal reconstruction pipelines that embed known structure while retaining flexibility through learning
- Low-latency and real-time inference for high-throughput experimental or sensing systems
- Sequential decision making and reinforcement learning under uncertainty and partial observability
I am especially drawn to approaches that preserve physical structure, remain numerically stable, and can be deployed and maintained in operational environments.
Contact
Email is the best way to reach me:
dowlingwong@gmail.com
You can also find research code and side projects on GitHub, and a more formal summary of my background in my CV and LinkedIn profile.