Research agenda
Applied machine learning, read closely
My research sits at the intersection of applied machine learning and algorithmic efficiency. The work below follows three threads: how models behave under constraint, how methods can be checked against honest benchmarks, and how the resulting code and data can be reproduced by others.
Publications
Papers and preprints
Peer-reviewed papers and preprints on applied machine learning and algorithmic efficiency. Each entry lists its co-authors, a short abstract, and a direct link to the open access PDF.
Peer-reviewed paper
Efficient inference under memory constraints
With co-authors. Abstract: a study of how quantization and pruning trade accuracy for latency on constrained hardware, with benchmarks across standard tasks. Open access PDF available.
Preprint
Sample-efficient evaluation of learned estimators
With co-authors. Abstract: an empirical comparison of estimator variance across training budgets, with a reproducible evaluation protocol. Preprint PDF available.
Technical monograph
Algorithmic foundations of efficient learning
A longform monograph on the mathematical formulations behind efficient learning, with worked examples and citation metadata for each chapter.
Research notes
Open questions and working hypotheses
The central question is how much accuracy a learned method can give up before its behavior changes qualitatively. My working hypothesis is that the threshold is predictable from a small set of measurable statistics, which would let practitioners choose a compression level without repeated full-scale runs.
This is an open problem. The experiments so far are limited to a few architectures, so the claim is provisional until the methodology is validated on a wider set of benchmarks.
Each investigation starts from a stated hypothesis, a fixed evaluation protocol, and a list of assumptions I expect to break. Negative results are recorded alongside positive ones, and the code used to produce each figure is published with the paper.