Beyond SHAP: Diagnosing Vector Embeddings with Visual Explainable AI
Speaker
Valdas Druskinis
Valdas is a Machine Learning Engineer at carVertical, where he builds reliable and interpretable ML systems with a strong focus on end-to-end model and data lifecycles. Previously, he has lived and worked in multiple countries, including time at Mercedes-Benz AG, and has built production systems across classical ML, deep learning, and GenAI.
Abstract
When your embedding-based classification model fails, should you collect more data or try a different approach? This talk shares a practical XAI workflow using UMAP visualization and prototype analysis to uncover systematic failures. We will explore how to use these tools to identify semantic overlaps and make evidence-based decisions when debugging high-dimensional similarity systems.
Description
When an embedding-based classification model fails, how do you know whether to collect more data, adjust the labeling strategy, or fine-tune the model? Traditional XAI tools like SHAP don’t help here—the real insight lies in the geometry of the embedding space.
This talk presents a compact, practical workflow for visually diagnosing vector embeddings using UMAP projections and prototype analysis. We’ll compare different projection methods, visualize data points, and show how to spot common failure modes such as semantic overlap, out-of-distribution inputs, and cluster drift after fine-tuning.
Attendees will learn a fast, evidence-based approach for debugging high-dimensional similarity systems—and for making smarter decisions about data, labeling, and model updates.