How internal neurons represent the short context: an emergent perspective

Natural language acquisition is a crucial research domain of artificial intelligence. To enable the computer acquire and understand the human language and generate the corresponding response, we must research the principle of language acquiring of human being. Most of the prior language acquisition methods use handcrafted internal representation which is not sufficiently brain-based. An emergent developmental network (DN) is presented to acquire the certain speech extracted by mel-frequency cepstrum coefficient (MFCC), from sensory and motor experience. This work is different in the sense that we focused on mechanisms that enable a system to develop its emergent representations from its operational experience. In this work, internal unsupervised neurons of the DN are used to represent the short contexts, and the competitions among the internal neurons enable them to represent different short contexts. To demonstrate the acquisition effect, we study and analyze the influences of different network structure (i.e., different neuron number and weight threshold) on the language acquisition rate. Four speech acquisition experiments demonstrate efficiently how such internal neurons represent the short context while they are not directly supervised by the external environment. The presented network is developmental which means that the internal representations are directly learned from the signals of the input and motor ports, not designed internally for particular task, hence the same learning principles are potentially suitable for other sensory modalities.