Yang Hu

First-author on Perceptual Initialization, raised zero-shot accuracy +7 pp / retrieval +45 % by injecting human perceptual priors into CLIP before large scale web pretraining. Boosted multimodal model human alignment 95 % (r 0.45→0.88) by PEFT on 4.7 M large scale image triplets and raised cross modality correlation up to +30 %. Accelerated evaluation pipeline for optimal model–human correlation by replacing traditional Bayesian optimization with a CUDA-enabled JAX-powered evolutionary algorithm for nonlinear search, achieving a 3–4× speedup.

Fine-tuned OpenAI Whisper for timeseries anomaly detection with PEFT, delivering ROC-AUC 0.95 for signal vs. noise and 95 % accuracy across 9 anomaly classes while cutting training compute by 99 % on a single GPU. Deployed real-time inference pipeline that outperformed legacy CPU stack with 4x faster convergence, recovered 80 % of online production events and filtered 98 % false positives, enabling sub-second alerts for downstream systems.

Developed, trained, and deployed machine learning model for ATLAS Experiment Pion energy reconstruction project. Applied Deep Sets framework and Mixture Density network to tackle variable length unordered set of particles data in Pion energy reconstruction. Helped develop a new course, Neural Network in Engineering and Physical Science, for the Physics department.

Relevant Coursework: Java Computer programming, Python Data programming, Statistical Methods, Neural Network in Engineering and Physical Science, Linear Algebra, Mathematical Physics, Statistical Physics, Visualization and Computer-Aided Design