Leah P.

Responsibilities:

• Designed a stable 0–1 scoring function that can wrap any regression model, so faithfulness metrics produce accurate and comparable scores across different models.

• Implemented a consistent input-processing layer supporting PyTorch, TensorFlow, YOLO, and ONNX models, ensuring repeatable and framework-agnostic explainability results.

• Designed a wrapper that converts multi-channel mask outputs into a single differentiable scalar, allowing segmentation models to be explained using gradient-based methods.

• Exposed and wired key explainer and metric parameters so users can highlight and tune the most influential settings directly from the application UI.

• Developed a custom pixel-removal pipeline for 2-channel SEM images, iterating through multiple experiments and refinements until achieving a stable faithfulness evaluation

• Led the research and integration of Quantus faithfulness metrics into the application, enabling reliable quantitative evaluation of explanation quality across a wide range of computer vision models (both Applied and external).

• After exploring alternative designs, implemented a loading mode that preserves gradient flow and exposes selected internal layers, enabling gradient-based explainability without changing model code.

• Built GUI components and a heatmaps library that display original images, overlays, and metric tables, and allow users to save and revisit multiple explanation runs per model.

• Designed and implemented a dynamic form system that fetches explainer and metric metadata from the server and sends back validated, structured configurations, keeping the frontend and backend loosely coupled.

• Designed and ran experiments comparing several explainability methods on Applied’s regression and segmentation models, analyzing their behavior and summarizing insights for the team.