SPIDER: Subspace Primitives that are Interpretable and Diverse
Abstract
Contribute three classes of primitives to the D3M program that are not only innovative machine learning methods but also provide a certain transparent mechanism by which the domain expert, data science-novice can naturally incorporate their expert knowledge to yield a better model. The three classes of primitives each attack the discovery of subspaces within the input data space with different tools: (1) multimodal embeddings leverage sparse models, (2) invariance discovery primitives, and (3) subspace clustering primitives.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 2021
- Accession Number
- AD1130271
Entities
People
- Jason J. Corso
- Laura Balzano
Organizations
- University of Michigan