Towards Explainable, Interactive, and Scalable Projection Techniques

May 13, 2021
2:00-2:40 pm ET
Sococo VH 302, Zoom
Speaker: Gabriel Appleby
Host: Remco Chang

Abstract

Quals talk:

Projection algorithms such as t-SNE or UMAP embed high-dimensional data in a low-dimensional space while preserving some notion of similarity. These methods are useful for visualization but can be difficult to tune and interpret. Many projection techniques depend heavily on important hyperparameters, that greatly affect the final visualization. Unfortunately, iteratively recomputing projections is computationally intensive and interaction patterns that support real- time adjustments are infeasible on current hardware. In this talk, I describe a scalable method that allows for real-time interactive hyperparameter tuning. This allows practitioners to easily embed scalable interactive projection in their visualization system. Even with a properly tuned projection, it can be difficult for non-expert practitioners to correctly interpret the visualization. To that end, I present a technique for fast inverse-projection that allows the practitioner to explore the space in real-time. This technique can also be used to produce powerful backgrounds for projection plots that aid understanding further, such as error maps and gradient maps. These background visualizations help the practitioner to understand how well they can trust the underlying projection and also give an indication of how distance in the low-dimensional space maps to distance in the original high-dimensional space.

Please join the meeting in Sococo VH 302, or Zoom.

Join Zoom Meeting: https://tufts.zoom.us/j/98610939077

PASSWORD: See colloquia email

Dial by your location: +1 646 558 8656 US (New York)

Meeting ID: 986 1093 9077

PASSCODE: See colloquia email