Self-Organizing Feature Maps and Their Application to Digital Coding of Information

Digital coding of information tries to compress information to improve the use rates of existing transmission and storage devices without degrading the quality after the decoding process. Vector Quantization (VQ) has been broadly used in such systems —speech and image coding for example— to increase the compression under some distortion constraints. Although some algorithms have been proposed for VQ design, all of them, either behave poorly when used through noisy transmission channels, or introduce a decission delay which avoids their use in many applications. Self-Organizing Feature Maps (SOFM) are a Neural Net architecture proposed by T. Kohonen for classification purposes. This paper presents an algorithm to adapt SOFMs for VQ problems, and studies two main advantages that this new structure offers: robustness against line errors and even higher information compression.