BYU

Abstract by Robert Pottorff

Personal Infomation


Presenter's Name

Robert Pottorff

Co-Presenters

None

Degree Level

Masters

Co-Authors

None

Abstract Infomation


Department

Computer Science

Faculty Advisor

David Wingate

Title

Learning Object-oriented Embeddings

Abstract

Learning to interpret pixels in terms of discrete objects would empower learning algorithms to dramatically reduce the state space over which they operate. My thesis proposes a neural network architecture which learns to decompose and compose scenes in terms of independent objects and demonstrates the usefulness of these independent representations on improving sample complexity in neural networks.