Fig. 3: Functional subpopulations in the STG correlate with different contextual representation layers in DNNs.
From: Dissecting neural computations in the human auditory pathway using deep neural networks for speech
a, Anatomical locations of all speech-responsive electrodes, mapped onto a common cortical space in the enlarged image of the boxed region. Different colors indicate different functional clusters. b, Averaged event-related potential (ERP) of each functional cluster. All time points were aligned with sentence onsets and normalized to the resting-state baseline (mean ± s.e.m.). c, Normalized BPSs of the encoding models based on every single layer in HuBERT for each functional cluster (maximum over delay window lengths). Red star indicates the layer with the highest score; black dot indicates other layers that were not statistically different from the best layer (P > 0.05, paired t test, two-sided; n = 83 electrodes for cluster 1, n = 61 electrodes for cluster 2). Box plot shows the first and third quantiles across electrodes (orange line indicates the median; black line indicates the mean value; and whiskers indicate the 5th and 95th percentiles). Horizontal gray line: the performance of the full acoustic-phonetic feature baseline model. d, Histogram of the optimal delay windows corresponding to models in c.
