Exploring hidden state dynamics in decision making

Project Description

The aim of the project is to explore the effectiveness of Hidden Markov Models in describing the neural dynamics.

Objectives and Approach

Decision making often involves recognizing cues, storing them in memory and performing an action based on the stimulus, stored memory and context. Hidden Markov Models (HMMs) are a class of models that describe observable events as the output of a set of hidden states and transition between states. We reasoned that animal behavior, and the underlying neural dynamics, could be described by HMMs.

Progress

• Preliminary exploration of the data (Done)
• Unsupervised clustering of the data (Done)
• Investigate why unsupervised clustering failed using pca decomposition (Done)
• Investigated with supervised dPCA to cluster task condition (Done)
• Building of HMM frameworks (Done)
• Fitting HMMs to the data (Done)
• Comparison and reliability of the fit HMM model states for mouse and monkey datasets (In progress)

Next Steps

• Investigate and implement approaches to remove the first principal component from the data and rerun analyses.
• Identify the percentage of neurons that are showing visual and memory related states.
• Testing alignments with respect to different task conditions to measure reliability and reduce noise in fits.
• If the identification of hidden states fails after removing the first principal component, are there other independent approaches we can try before concluding our assumptions are wrong?