Figure 5. Adaptation of spike timing through measured changes in gain and spike-threshold. A, Cross-validated log-likelihoods (LLx) of the LN (red), GL (blue), and GN (green) models on successive repeated trials of the same stimulus, demonstrating that the first trials of the GNM are very poorly predicted, but this quickly changes and stabilizes over several trials. bits/spk, Bits per spike. B, Top, Gains of the excitatory (αE), suppressive (αS), and spike history (αRP) terms of the GN model that best predict the spike train of the first 5 s of each trial. Bottom, The simultaneously fit offset Δ of the spike threshold. Together, this suggests that the gains of the excitatory and suppressive terms are initially much smaller at the beginning of the trial, but the neuron is closer to spike threshold, and these values stabilize over time. C, Applying these gains and offsets dramatically stabilizes the quality of the prediction of the GNM, as demonstrated by the LLx of the second 5 s of the repeated stimulus for the GNM (green). Similar adjustments to the gains and offsets of the other models (LN, red; GL, blue) have little effect on the LLx. D, The difference in predictions of the GNM from the first trial (cyan) and at stabilized values (black) predict that spike times in general should be earlier initially. This agrees with observed spike trains on the corresponding section of cross-validation stimulus (top), which shows the observed spike trains of the first trial (cyan) compared with later trials (black). The small differences in spike timing between the first and later trials explain why this was only detected in the LLx of the GNM (A).