accuracy of neuronal encoding depends on the response statistics of individual neurons and the correlation of the activity between different neurons. Here, the dynamics of the neuronal response statistics in the anterior superior temporal sulcus of the macaque monkey is described. A transient reduction in the normalized trial-by-trial variability and decorrelation of the responses with both the activity of other neurons and previous activity of the same neuron are found at response onset. The variability of neuronal activity and its correlation structure return to the levels observed in the resting state 50 -100 ms after response onset, except for marked increases in the signal correlation between neurons. The transient changes in the response statistics are seen even if there is little or no stimulus-elicited activity, indicating the effect is due to network properties rather than to activity changes per se. Modeling also indicates that the observed variations in response variability and correlation structure of the neuronal activity over time cannot be attributed to changes in firing rate. However, a reset of the underlying spike-generating process, possibly due to the driving input changing from recurrent to feedforward inputs, captures most of the observed changes. The nonstationarity indicated by the changes in correlation structure around response onset increases coding efficiency: compared with the mutual information calculated without regard to the transitory changes, the decorrelation increases the information conveyed by the initial response of modeled neuronal pairs by Յ24% and suggests that an integration time of as little as 50 ms is sufficient to extract 95% the available information during the initial response period.
I N T R O D U C T I O NThe ability of the brain to encode information is determined by the response characteristics of the individual neurons and the correlation structure between the responses of different neurons. Reports of the neuronal coding of visual stimuli have typically analyzed data using sample windows that are fixed in both time and size (e.g., a window size of 250 ms starting 50 ms after stimulus onset). More recently, studies have started to examine changes in the response characteristics at different time points within the stimulus-elicited response (e.g., Amarasingham et al. 2006;Churchland et al. 2006;Muller et al. 2001;Smith and Kohn 2008). The study of the dynamics of response statistics can help elucidate the properties of underlying neural circuits and constrain computational models in addition to quantifying how dynamic changes in response characteristics influence the stimulus-related information carried by the responses.To determine the impact of changes in response statistics over time in terms of information, it is necessary to measure response variability. For individual neurons, the more variable the responses to a given stimulus, the less information those neurons can encode. The trial-by-trial variability increases from retina to the lateral geniculate n...