Figure 3. Methods for predicting spiking activity.

(A) Illustration of “2D STA” method. The amplitude and derivative of the envelope is computed at a fixed time lag before the occurrence of each spike (here 15 ms). The mean firing rate as a function of amplitude and derivative can then be computed by a smoothing method previously used to compute hippocampal place fields.

(B) Prediction from the 2D STA depends on time lag. Firing rate maps in amplitude/derivative space (2D STAs) are computed from a training set, for multiple time lags (cyan: 18 ms; green: 15 ms; magenta: 12 ms). To predict firing rate on the test set, the amplitude and derivative of the envelope, lagged appropriately, are used to index into the 2D STA. The quality of prediction is evaluated on the test set using the log-likelihood ratio of the observed spike train under the prediction, compared to the cell’s mean firing rate. The result (lower-right) is an estimate of prediction quality which varies as a function of time lag.