DESIGN OF NEURAL NETWORKS FOR LOSSLESS DATA-COMPRESSION

The author describes a novel design of neural networks for lossless da ta compression. The proposed neural network establishes an efficient d ictionary by storing two symbols into each neuron and interconnecting those neurons that match a number of consecutive input strings. After an operation cf experience-based competitive learning, a number of inp ut strings can be matched by winning neurons in the network. Variable- length codes are then designed to encode the location of the first neu ron, possible interconnections, and the number of matched neurons to a chieve data compression. For unsuccessfully matched input strings, a l iteral code is constructed that contains an overhead code identifying the length of the literal code and the original codes of those unsucce ssful strings. Extensive experiments show that the proposed neural net work achieves very competitive compression performance in comparison w ith a few typical existing data compression algorithms. This also open s a new area for the application of neural networks to lossless data c ompression, where massive parallel processing and powerful learning ca pability can be utilized to develop high-performance algorithms and ne w techniques. (C) 1996 Society of Photo-Optical Instrumentation Engine ers.