Latent variable sequence identification for cognitive models with neural network estimators.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Extracting time-varying latent variables is crucial for understanding dynamic cognitive processes.
• Current methods are limited to specific cognitive models, excluding those with intractable likelihoods.

Purpose of the Study:

• To develop a simulation-based approach using recurrent neural networks (RNNs) for inferring latent variable sequences.
• To overcome limitations of existing methods for cognitive models with intractable likelihoods.
• To enable broader exploration of computational cognitive models.

Main Methods:

• A simulation-based approach leveraging recurrent neural networks (RNNs).
• Mapping experimental data directly to the latent variable space.
• Utilizing simulated data for training and validation.

Main Results:

• Achieved competitive performance in inferring latent variable sequences for both likelihood-tractable and intractable models in simulations.
• Demonstrated applicability on real-world datasets.
• The approach is practical for individual data, generalizable, and adaptable to continuous/discrete latent spaces.

Conclusions:

• The proposed method expands the range of cognitive models researchers can analyze.
• Facilitates testing a wider array of cognitive theories by enabling inference in complex models.
• Combines RNNs and simulated data for robust latent variable extraction.