The Japanese journal of neuropsychology

Full Text of this Article
in Japanese PDF (120K)

ArticleTitle	A Two-Staged Neural Network Model for Letter Identification and Its Application to the Comparison between Disconnection and Visually Impaired Hypotheses
Language	J
AuthorList	Shinichi Asakawa
Affiliation	Centre for Information Sciences, Tokyo Woman's Christian University
Publication	Japanese Journal of Neuropsychology: 18 (2), 92-100, 2002
Received	Oct 12, 2000
Accepted	Dec 8, 2001
Abstract	This study intended to explain the computational role of two brain areas (posterior fusiform gyrus and anglar gyrus) playing in reading behavior of Japanese characters, and to show the computer simulation of impairments when these areas were damaged. The model proposed is consisted of two stages. In view of the large number of Japanese characters (100 times more than in European languages), it is impractical for both human and computer to adopt a simple layered network model like perceptron to Japanese characters, although this model generally applied to the European letter recognition. At the first stage, self organizing feature mapping (Kohonen, 1985), visual inputs are roughly and rapidly classified into groups on the basis of similarity of shape. The grouping at the first stage can be regarded as a representation of the topological organization frequently observed in the various areas of the brain. The result of the first stage is forwarded to the second stage (a backpropagation network). The role of this stage is to identify the character among the candidates which are the product of the first stage, and to give an abstract and modality free expression of input for further processing systems. The model can be regarded as an implementation of the roles that two different areas in the brain. Gakushu kanji set consisted of 1006 characters were rasterized to 32× 32 pixels as input images. A Gaussian and a canonicalization operators were applied for pre-processing images. The 16-bits length of JIS X 0208 kanji code were employed as teacher signals. After training, the system showed 89% correct ratio of identification. Two numerical experiments with disconnection and destruction were performed to demonstrate two kinds of neuropsychological hypotheses, visuo-speech disconnection (Geschwind, 1965) and visually-impaired (Farah, 1990) hypotheses. According to disconnection hypothesis, some units in the hidden layer at the second stage ought to be removed. On the other hand, some amounts of noise should be added to the weight vectors at the first stage in order to simulate the visually-impaired hypothesis. Damage in the second stage or the disconnection hypothesis causes severe lesion, in fact the system's performance decreased dramatically as a function of severity of damage. On the contrary, the accuracy was not so degraded in case of damage in accordance with the visually-impaired hypothesis. Thus, symptoms predicted by these two hypotheses might be differ each other.
Keywords	morphological dyslexia, neural network, visuo-speach disconnection hypothesis, visually-inpaired hypothesis