When representing speech signals as groups of spatio-temporal events, a problem is to find and group the events. Here, we propose to do so using self-organising cable neuron networks as found in the theory of Neuronal Group Selection. The basic principle of the learning rules that establish the process of group forming is reinforcement of coincidences: by differences in transmission delay times and post-synaptic integration times, a certain spatio-temporal input pattern will give rise to more or less simultaneously activation of several neurons. These neurons will tend to amplify the interconnections inbetween them and weaken their connections with other neurons. Once having identified groups of spatio-temporally correlated events at the lowest level, the grouping process can be repeated at the following level, i.e. grouping these groups, in order to recognise higher level entities.
Bibliographic reference. Klaassen, Arno J. (1993): "Grouping of acoustical events using cable neurons and the theory of neuronal group selection", In EUROSPEECH'93, 381-384.