Active Learning for Discrete Latent Variable Models.

Area of Science:

• Machine Learning
• Computational Neuroscience
• Statistical Modeling

Background:

• Active learning is crucial for efficient model training.
• Latent variable models are vital in neuroscience and psychology.
• Previous active learning methods overlooked latent variable models.

Purpose of the Study:

• Propose a novel framework for maximum-mutual-information input selection for discrete latent variable regression models.
• Address the gap in active learning research concerning latent variable models.
• Demonstrate the efficacy of active learning for complex models.

Main Methods:

• Developed a maximum-mutual-information input selection framework.
• Applied the method to mixtures of linear regressions (MLR) and generalized linear model (GLM) hidden Markov models (HMMs).
• Utilized Fisher information for analytical insights and validated with simulations and real-world data.

Main Results:

• Active learning provides significant gains for mixtures of linear regressions, unlike simple linear-gaussian models.
• The proposed method substantially reduces data requirements for fitting GLM-HMMs.
• Outperformed variational and amortized inference methods in fitting GLM-HMMs.

Conclusions:

• Maximum-mutual-information input selection is effective for latent variable models.
• This approach offers a powerful way to characterize temporally structured latent states.
• Applications span neuroscience, psychology, and various scientific disciplines.