The raw waveforms were separated into field potentials and spike trains using low- and high-pass filtering (for field potential: second-order Butterworth low-pass filter, 20-Hz cutoff frequency; for spikes: second-order high-pass filter, 100-Hz cutoff frequency).
Because the recordings within approximately 5 ms of the stimulus onset were usually obscured by stimulation artifacts, an artifact removal technique based on average subtraction was applied to recover early-phase (short-latency) RGC spike responses.
21,22 Two distinct classes of recorded waveforms near the stimulus onset, with or without evoked RGC action potentials, could be visually identified. The average waveforms of the stimulus artifact could be estimated from multiple waveforms that did not contain the evoked action potentials. The RGC spike waveforms within approximately 5 ms of the stimulus onset could be recovered by subtracting the averaged stimulus artifact obtained from single trial waveforms containing action potentials.
Spike sorting software (Offline Sorter; Plexon Inc, Dallas, TX) was then used to transform the waveforms containing multiunit activities into multiple single-unit spike trains. The number of poststimulus RGC spikes was counted to measure the RGC response strength and to analyze the modulation of the RGC response strength by pulse amplitude. To quantify the accuracy of the input information represented by the RGC responses, similarities between the pulse amplitude variation time series and the RGC response strength time series were computed by coding fraction.
23 The coding fraction, γ, is defined as follows:
where ε and σ
s denote the root mean square error of estimating pulse amplitude time series from firing rate time series and standard deviation of pulse amplitude time series, respectively; ε can be calculated from the difference between pulse amplitude time series and firing rate time series. When the firing rate is not at all correlated to the input (i.e., pulse amplitude), the mean square error becomes equal to the variance of stimulus σ
s 2 and, thus, γ becomes 0. Conversely, when the two time series are perfectly correlated, ε is zero and, thus, γ becomes 1. Hence, the coding fraction provides a convenient method to quantify the accuracy of stimulus encoding.
23 We evaluated the accuracy of temporal information on the input that is encoded in RGC responses by the coding fraction.