/venditto_glm-hmm

Primary LanguageMATLAB

GLM-HMM for binary observations

MATLAB functions for fitting a GLM-HMM on behavioral data in tasks with a binary choice. Implementation based on Bishop's "Pattern Recognition and Machine Learning" and Escola et al. (2011), Neural Computation. To use, include both fitGlmHmm.m and runBaumWelch.m need to be in your MATLAB path.

Model description

Coming soon

Included functions

See function description and additional parameter descriptions by inputting help fitGlmHmm or help runBaumWelch into MATLAB command window.

fitGlmHmm: Fitting the model

[model, ll] = fitGlmHmm(y,x,w0)

Required inputs:

  • y: (1 x NTrials) binary observation data
  • x: (NFeatures x NTrials) design matrix; behavioral features used to predict observation data
  • w0: (NFeatures x NStates) initial latent state GLM weights. Desired number of latent states is taken implicitly from the second dimension

Outputs:

  • model: struct containing fit parameters
    • w: (NFeatures x NStates) latent state GLM weights
    • pi: (NStates x 1) initial latent state probability
    • A: (NStates x Nstates)
  • ll: (1 x NIter) log-likelihood of model fit at each iteration

runBaumWelch: Computing latent state probabilities or fit likelihood on some data set

[gammas,xis,ll] = runBaumWelch(y,x,model)

Required inputs:

  • y and x as described above
  • model: output of runGlmHmm.m

Outputs:

  • gammas: probability of latent state given model parameters for each trial
  • xis: joint posterior distribution (summed across trials). can be used to calculate the estimated transition matrix
  • ll: log-likelihood of the model

Example

The MATLAB script example_glmhmm_fit.m details the process of fitting a GLM-HMM on an evidence accumulation ('accumulating towers') task data. This script:

  1. Simulates a GLM-HMM model to generate behavioral data
  2. Initializes necessary input and fits a GLM-HMM model to generated data (i.e. recovers GLM-HMM used to generate data)
  3. Compares recovered fit to simulated model

Credit

Written by Sarah Jo Venditto. Please cite this repository