Efficient Visual Appearance Optimization by Learning from Prior Preferences
summary
Meta-PO integrates preferential Bayesian optimization with meta-learning, stores prior users' surrogate models and uses similarity-weighted acquisition functions plus a two-step acquisition to propose a 2D gallery centered on the current best design. The system applies a time-decay to prior weights, samples images on the plan, and updates the current model from pairwise preferences. In a 2D/3D appearance study baseline needs 9.5 iterations, Meta-PO converges in ≈5.86 iterations for shared themes and ≈8 iterations across divergent themes.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:24Current methods for photo adjustments
- 1:27Limitations of preferential Bayesian optimization
- 2:09Crash course on Bayesian optimization
- 3:14Challenges in human preference optimization
- 3:36Introducing MetaBO method
- 4:40How MetaBO selects candidates
- 6:25User study design and conditions
- 7:50Results of the user study
- 8:52Generalizing MetaBO to other tasks
- 9:14Conclusion and internship search
- 9:37Q&A session
related talks
38mPractical Few-Shot Learning
26mAddressing Some of Challenges When Optimizing Long-to-Train-Models
32mImprove LLMs Alignment with Complete and Robust Preference Data
14mUnconference: Self-optimizing deep research agents
10mSemanticAdapt: Optimization-based Adaptation of Mixed Reality Layouts Leveraging ...
33mAutomatic Agent Optimization and the Rise of Self-Improving Agents
33m"Are You Sure?": blending product comparisons and recommendations with A.I.
30mIs Prompt Engineering Dead?