Zhengjie (Jane) Yang

CS231n · Deep Learning for Computer Vision · 2026

Improving Bird Image Classification by Combining Visual and Language-Based Learning Approaches

Jane Yang, Marco Vizcarra, and Adem Rimpapa

A team study of fine-grained recognition across 555 NABirds classes, combining a part-aware ConvNeXt model with SigLIP vision-language representations through late fusion. Experiments compared text prompts and examined errors between visually similar bird categories.

The fused model achieved 92.48% top-1 accuracy in the reported test evaluation, compared with 91.00% for the ConvNeXt model alone.

CS224R · Deep Reinforcement Learning · Spring 2026

Diversity-Aware RLOO for Pass@k Reasoning in Countdown

Jane Yang and Will Nathaniel Hansen

Studied reward shaping for Qwen2.5-0.5B that rewards distinct, verified solutions to arithmetic problems. I implemented the Modal training and evaluation workflow and diversity-aware reward pipeline, ran SFT/IPO/RLOO experiments, and analyzed solution coverage.

The best diversity-aware run improved pass@16 from 0.72 to 0.78 over vanilla RLOO on 50 held-out prompts. IPO reached 0.80; the study did not include multiple training seeds.

ENERGY SYSTEMS · 4D DESIGN · MODULAR CONSTRUCTION

PowerTwin 4D: Modular Data Center Design and Fast Delivery

An independent prototype exploring how electricity supply, grid interconnection, and energy-storage choices can inform modular data-center design and construction sequencing. An interactive 3D model and 4D timeline connect power configurations, equipment dependencies, and site deployment.

I use this project to explore the link between energy-system planning and time-to-power. The demo uses illustrative planning assumptions; its outputs are not validated engineering designs.