High-level interface

To estimate a sparse variational Gaussian process factor analysis model we:

1. Construct an empty model

   model = stats.svGPFA.svGPFAModelFactory.SVGPFAModelFactory.buildModel(
       conditionalDist=stats.svGPFA.svGPFAModelFactory.PointProcess,
       linkFunction=stats.svGPFA.svGPFAModelFactory.ExponentialLink,
       embeddingType=stats.svGPFA.svGPFAModelFactory.LinearEmbedding,
       kernels=kernels)
   

   by specifying a conditional distribution (e.g., point-process, PointProcess), an embedding type (e.g., linear, LinearEmbedding), a link function (e.g., exp()) and providing a set of kernels (Kernel).

2. Estimate the parameters of the model

   svEM = stats.svGPFA.svEM.SVEM()
   lowerBoundHist = svEM.maximize(model=model, measurements=spikeTimes,
                                  initialParams=initialParams,
                                  quadParams=quadParams,
                                  optimParams=optimParams)
   

   by providing a set of measurements, spikeTimes, initial parameters, initialParams, quadrature parameters, quadParams and optimisation parameters, optimParams.

3. Plot the lower bound history of the estimated model

   plot.svGPFA.plotUtils.plotLowerBoundHist(lowerBoundHist=lowerBoundHist)
   

   

and model parameters (e.g., latents).

plot.svGPFA.plotUtils.plotTrueAndEstimatedLatents(times=testTimes, muK=testMuK, varK=testVarK, indPointsLocs=indPointsLocs, trueLatents=trueLatentsSamples, trialToPlot=trialToPlot)

Please refer to the following notebook for a full piece of code running svGPFA and plotting its estimates.