SupplementaryMaterials: ProbabilisticTensor DecompositionofNeuralPopulationSpikingActivity

UpdatesforoffsettensorV For d = (d 1,...,d D1), we want to updateν d . Thisprovides justification for the approximation, the empirical quality of which we then evaluate in numerical experiments. Its value arises solely in inference, where the conditional likelihood onv, that isp(x|v,u) viewed as a function ofv, is simplified to a Gaussian-conjugate form. The full VB model affects this limit in two ways. However, a similar limit applies if the model provides a good fit to the data with low dispersion.