skip to primary navigation skip to content

CBSU bibliography search

To request a reprint of a CBSU publication, please click here to send us an email (reprints may not be available for all publications)

A neuronal model of the language cortex
15th Annual Computational Neuroscience Meeting
Year of publication:
CBU number:
During speech production, two different areas of the brain become simultaneously active: primary motor cortex (responsible for sound articulation) and auditory cortex (stimulated by the produced sounds). We postulate that during language acquisition, the repeated simultaneous activation of these two areas leads, through Hebbian learning, to the formation of strongly-associated cell assemblies, which are: (a) distributed across cortical areas, (b) word-specific, and (c) activated even by only partial (e.g., auditory) stimulation. In order to test the validity of this hypothesis, we: (1) simulated learning of neural assemblies in a biologically-motivated computational model of the left perisylvian language cortex; (2) observed the functional characteristics of the resulting model, and (3) explored its neurophysiological properties and compared them with real experimental data. Six cortical areas were simulated: (i) primary auditory cortex, (ii) auditory belt and (iii) parabelt areas (Wernicke areas), (iv) inferior prefrontal and (v) premotor cortex (Broca areas), and (vi) primary motor cortex. Each area (layer) consisted of a 25x25 grid of graded-response neurons. The network included random recurrent, forward and backward links; recurrent connection probabilities decreased with increasing cortical distance following a Gaussian distribution; between-area links were topographic and followed the same probability distribution; non-adjacent layers had no connections. Local and global inhibition mechanisms controlled activity within each layer. The model was confronted with simultaneous patterns of activations in both its auditory (1) and motor (6) layers, as one would expect it in early speech production. During training, the synaptic weights of all excitatory links were modified according to the co-variance learning rule [Sejnowski, 1977]. We observed formation of cell assemblies, different assemblies responding selectively to different input patterns. We also observed the presence of spontaneous rhythmic activity in the network. During the testing phase, the following functional characteristics emerged: when only the auditory layer was stimulated with one of the learnt patterns (words), we observed that (a) the relevant assembly was immediately and strongly activated in that layer and started to oscillate; (b) the rhythmic activity of the assembly in that layer spread to the remaining layers; (c) presentation of new, previously unseen "pseudo-patterns" (pseudo-words) produced, on average, a smaller response in the network than that produced by words. These results are consistent with recent evidence obtained from EEG recordings of the brain activity during speech-listening tasks, and help explain neurophysiological observations about cortical processing of words and pseudo-words in terms of a clearly-defined neurobiological model of language.