The contribution of trained parameters to the goodness of fit of a Bayesian neural encoding model for the auditory system

A Bayesian neural decoding model requires the use of a neural encoding model. The parameters of this encoding model are generally fitted to some training data and used subsequently in the decoding of a test stimulus. One encoding model that has been widely used for the auditory system implies the use of a generalized linear model (GLM) having three parameters accounting respectively for the spontaneous rate of the neuron, its spectro-temporal receptive field and the dynamics of the neuron. Here we present a cross-validation study of the goodness of fit of a GLM encoding model in order to quantify the effects on the fitting of using model parameters estimated from a training data set on a test data set. The goodness of fit is measured using Kolmogorov-Smirnov (KS) statistics. It is observed that using trained parameters on the test data yields a much poorer than expected goodness of fit of the auditory encoding model, with only 5% of the neurons having a suitable fit. Moreover, we show that this poor goodness of fit is the result of all three parameters of the auditory GLM encoding model being inadequate for the test data. Using such parameters in a decoding framework may thus result in a biased decoded stimulus.