< Back to Article

Unsupervised learning of temporal features for word categorization in a spiking neural network model of the auditory brain

Fig 6

Evidence for PGs responding selectively to word “one” or “two” in the A1 layer of the trained AN-CN-IC-CX model.

Each plot shows an example of a stable spatio-temporal spike pattern in A1 (red circles) in response to different pronunciations of the words “one” (left) and “two” (right). These spikes take part in at least one polychronous group that is selective for the particular word. In other words, these patterns are more likely to appear when an example of their preferred word is pronounced compared to an example of a non-preferred word. When projected through the A1→Belt connections () with different conduction delays (Δ_ij) (arrows), these patterns produce near-synchronous input from several A1 neurons onto a subset of Belt neurons (green, yellow or white circles corresponding to three separate Belt neurons with different distributions of axonal conduction delays Δ_ij). The green and yellow circles show such inputs for two Belt neurons which in this manner respond selectively for a number of different pronunciations of the word “one”, the white circles show inputs for a neuron that responded selectively to exemplars of the word “two”. Abscissa represents the time window Δt = t_j ± 50 ms around the origin. The origin is centered around all the times t when a chosen A1 neuron j fires (see Section Polychronization Index for details). Ordinate represents the 1000 neurons that make up A1 in the AN-CN-IC-CX model. Red circles show the ten elements of the firing pattern matrix with the largest mean spike counts (see Section Polychronization Index for details).

doi: https://doi.org/10.1371/journal.pone.0180174.g006