Hierarchical Latent Structure Learning through Online Inference

Area of Science:

• Cognitive Science
• Machine Learning
• Computational Neuroscience

Background:

• Learning systems require representations balancing generalization and discrimination.
• Existing models like online latent-cause models assume flat structures, while hierarchical Bayesian models require offline inference.
• There is a need for computational frameworks supporting online inference of hierarchical latent structures.

Purpose of the Study:

• Introduce the Hierarchical Online Learning of Multiscale Experience Structure (HOLMES) model.
• Develop a computational framework for hierarchical latent structure learning through online inference.
• Enable tractable trial-by-trial inference over hierarchical latent representations without explicit supervision.

Main Methods:

• Utilized a variation on the nested Chinese Restaurant Process prior.
• Employed sequential Monte Carlo inference for online processing.
• Developed a framework for unsupervised, hierarchical structure discovery in sequential data.

Main Results:

• HOLMES matched predictive performance of flat models with more compact representations.
• Demonstrated one-shot backward transfer to higher-level latent categories.
• Achieved above-chance outcome prediction in forward transfer tasks with novel feature combinations by exploiting abstract representations.

Conclusions:

• HOLMES provides a tractable computational framework for discovering hierarchical structure in sequential data.
• The model effectively balances generalization and discrimination for enhanced learning.
• HOLMES facilitates the learning of abstract representations crucial for transfer learning.