Shared Kernel Information Embedding for discriminative inference.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Latent variable models (LVMs) like GPLVM help prevent overfitting on small datasets.
• Existing LVMs face challenges including complexity, lack of explicit mappings, multimodality handling, and undefined latent space densities.

Purpose of the Study:

• To introduce novel LVMs, Kernel Information Embedding (KIE) and shared KIE (sKIE), that overcome limitations of current methods.
• To develop a model capable of handling multimodality and providing explicit latent space mappings.
• To enable robust inference with missing or partially labeled data.

Main Methods:

• Proposed Kernel Information Embedding (KIE) defining a joint density over input and latent spaces.
• Introduced shared KIE (sKIE) for modeling multiple input spaces with a single latent representation.
• Demonstrated quadratic learning, effective on small datasets, and suitability for online learning on large datasets.

Main Results:

• KIE and sKIE successfully define coherent joint densities and handle multimodality.
• The models permit inference with missing data and learning with partially labeled data.
• sKIE effectively facilitates learning local online models for large datasets, applied to human pose inference.

Conclusions:

• KIE and sKIE offer significant improvements over existing LVMs, particularly in handling complex data and missing information.
• These models provide a flexible framework for various machine learning tasks, including human pose estimation.
• The ability to learn local online models with sKIE opens possibilities for large-scale, dynamic data analysis.