Online Graph Completion: Multivariate Signal Recovery in Computer Vision.

Area of Science:

• Computer Vision
• Machine Learning
• Graph Theory

Background:

• Human-in-the-loop (HITL) systems integrate human supervision with machine learning algorithms for data acquisition and inference.
• Classical active learning methods are limited in addressing practical challenges like partial measurements and financial constraints in HITL paradigms.
• Sequential data acquisition from matrices or tensors necessitates novel completion and collaborative filtering strategies.

Purpose of the Study:

• To address the limitations of existing active learning methods in complex HITL scenarios.
• To develop a novel graph-based completion strategy for sequential data acquisition with human feedback.
• To improve the efficiency and accuracy of data annotation and inference in computer vision and other domains.

Main Methods:

• Formulated a graph completion problem for sequential measurement acquisition.
• Designed an optimization model in the Fourier domain of the graph.
• Utilized adaptive submodularity principles to develop practical algorithms.

Main Results:

• Demonstrated promising results on a large dataset of Imgur images, particularly for difficult-to-categorize images.
• Showcased the effectiveness of the proposed method in improving image annotation via crowdsourcing.
• Applied the approach to an experimental design problem in neuroimaging, yielding positive outcomes.

Conclusions:

• The proposed graph-based completion method effectively handles sequential data acquisition in HITL systems.
• Adaptive submodularity provides a robust framework for optimizing data acquisition strategies in machine learning.
• The approach has broad applicability in computer vision, crowdsourcing, and neuroimaging experimental design.