Charles Zhou

Explores representation learning on architectural ornament datasets, testing whether domain structure can replace large-scale pretraining. Finds that repeated local sampling enables training-from-scratch CNNs to match ImageNet-pretrained performance, emphasizing the value of curated design archives for ML.

A framework was developed to evaluate satellite clock stability using 1WCP and 3WCP carrier-phase methods. Validated on GPS data, the tool will be used to analyze timing performance of the upcoming NTS-3 mission.

An empirical study of failure modes in multi-agent debate for LLMs. We show that debate stability depends on the temporal structure of critique, with fine-grained, streaming interventions reducing hallucination cascades relative to batch-style debate.

This study investigates generalization failures in CNN-based chest X-ray classifiers. Models trained on the VinBigData datatset achieve >95% in-distribution accuracy but drop to ~26% on CheXpert, revealing the limits of transferability across medical imaging datasets.