Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Adaptation of Neural Network Model: Comparison of Multilayer Perceptron and LVQ

Dongxin Xu, Dao Wen Chen, Qian Ma, Bo Xu, Taiyi Huang

National Lab. of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, P.R.China

This paper explores the adaptation features of Multilayer Perceptron and LVQ3 under the speech phoneme recognition experiment. The result shows that Multilayer Perceptron can adapt to novel data easily but may lose the high performance on old data, and by contrast LVQ3 can maintain what it has learnt in the past but is hard to adjust itself to fit the novel data* if a relative great difference lies between the old and novel data. In order to gain a better adaptation algorithm, some modifications are added to LVQ3 and better results are achieved under the same experiment. A method to add in new code vectors for LVQ3 is also examined and a better result is also achieved. Finally, Some problems of the proposed algorithms are discussed.

Full Paper

Bibliographic reference.  Xu, Dongxin / Chen, Dao Wen / Ma, Qian / Xu, Bo / Huang, Taiyi (1994): "Adaptation of neural network model: comparison of multilayer perceptron and LVQ", In ICSLP-1994, 1567-1570.