Fig. 2: Accounting for non-stationarity in neural recordings.

a, Example neural activity in response to the same sound (speech at 60 dB SPL) presented at 5 different times within 2 individual recordings. Units were ordered on the basis of their CFs. Brighter colours indicate higher activity. b, The mean activity (per bin) of 20 randomly selected units from the non-stationary recording shown in a. c, The predicted neural activity for the non-stationary recording in a using the time-invariant and time-variant models. d, A schematic diagram of the time-variant DNN architecture. e, The median predictive power of the time-invariant and time-variant models across all units at five different times during the non-stationary recording shown in a. Predictive power was computed from the simulated and recorded neural responses as the fraction of explainable variance explained for a 30-s speech segment presented at 60 dB SPL. f, Performance comparison across nine animals and four sounds between time-variant and time-invariant DNN models. Each symbol represents the average across all time bins and units for one animal in response to one sound. g, Performance difference between models (time-variant − time-invariant) as a function of recording non-stationarity across seven animals and four sounds.