Infinite hidden conditional random fields for human behavior analysis.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Discriminative latent variable models like Hidden Conditional Random Fields (HCRFs) are effective for learning hidden structures in classification.
• Determining the optimal number of hidden states for HCRFs typically requires external validation, adding complexity and computational cost.

Purpose of the Study:

• Introduce the infinite Hidden Conditional Random Field (iHCRF), a nonparametric model designed to automatically learn the optimal number of hidden states for classification tasks.
• Demonstrate the iHCRF's ability to infer hyperparameters and converge to an accurate number of hidden states.

Main Methods:

• Utilized hierarchical Dirichlet processes to develop the nonparametric iHCRF model.
• Employed Markov-chain Monte Carlo (MCMC) sampling for efficient hyperparameter learning.
• Validated the model's performance on challenging classification tasks, including recognizing agreement, disagreement, and pain.

Main Results:

• The iHCRF model successfully converged to the correct number of represented hidden states.
• iHCRF significantly outperformed the best finite HCRFs, selected via cross-validation, on complex recognition tasks.
• Achieved superior performance with substantially reduced overall training, validation, and testing time.

Conclusions:

• The iHCRF offers an effective nonparametric solution for classification problems requiring latent structure discovery.
• This model automates the selection of hidden states, enhancing both accuracy and computational efficiency compared to traditional HCRFs.