A general model unifying the adaptive, transient and sustained properties of ON and OFF auditory neural responses
Fig 2
Overview of the proposed processing pipeline.
(A.) Its inputs are given by a spectrogram representation of the sound stimulus (e.g. obtained from a Gammatone filterbank). (B.) Two AdapTrans filters are convolved to each frequency band along the temporal dimension, effectively accounting for transient and permanent features of the stimulus. (C.) The filter outputs consist of a 2-channel, ON-OFF spectrogram that is further processed by conventional models as found in the litterature. (D.) Such models can be separated into several classes, depending on whether they are based on a single and large spectro-temporal receptive field (STRF), a cascade of convolutions with small-sized kernels (CNNs), or recurrent neural networks (RNNs)(the architectures of the models used in this study are shown in Fig 3). (E.) All models output the predicted neural activity as a time series that can be compared to a groundtruth recording.