45 min
Signed Rectified Flow: Negativity-Controlled Generation
Machine Learning, Artificial Intelligence and seminar by Qiang Liu, UT Austin
Wednesday 09:00 Los Angeles (GMT-7)
Recording available
Berkeley, United States
Recording
Abstract
Signed Rectified Flow incorporates both examples to favour and distributions to avoid. Although a signed measure cannot itself be sampled as a probability distribution, the construction induces a valid process that excludes negatively dominated regions. A continuity-equation analysis and charged-particle interpretation explain the barriers. Experiments explore how positive and negative information affect generation quality, memorization and exclusion constraints.
Topics
Signed Rectified Flowsigned measureprobability distributionnegatively dominated regionscharged-particle interpretationnegative informationgeneration qualitymemorization