Quantitative model comparison.

<p>Linear, quadratic and spike-triggered mixture models (STM) were evaluated on 8 slowly adapting cells (<b>A</b>) and 10 rapidly adapting cells (<b>B</b>). The performance of each model is measured in terms of the cross-entropy (negative log-likelihood) averaged over all cells (smaller is better). Light bars correspond to models which ignore the spike history, dark bars correspond to models which explicitly take the spike history into account. By subtracting the cross-entropy from the estimated entropy of the spike trains (“Prior”), an estimate of mutual information (MI) between stimuli and spike trains is obtained. The bars in <b>C</b> and <b>D</b> show (from left to right) the differences in performance between the linear model and the prior, the quadratic model and the linear model, and the STM and the quadratic model (with and without spike history dependency, respectively). The two right most bars show the improvement of the PSTH over the STM with and without spike history dependency.