Analysis-Driven Design of Representations For Sensing-Action Systems
Abstract
We have developed what is, to the best of our knowledge, the first complete theory of representation for decision and control task, which has shown not only to encompass and explain all known phenomenology in deep neural network-based representation learning, but also to predict phenomena that were thus far unexplained.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2017
- Accession Number
- AD1040878
Entities
People
- Stefano Soatto
Organizations
- University of California