Method and apparatus for encoding and decoding data utilizing data compression and neural networks

A data compression method which is put into effect with the aid of an encoder for transforming input data with the aid of a coding neural network provided at a transmission end of a channel, said coding neural network having a number (N) of input neurons greater than or equal to the number (P) of output neurons, and a decoder which comprises a decoding neural network which is matched to the coding neural network positioned at a receiving end of said channel for effecting the inverse transform of the encoder with a number (P) of input neurons less than or equal to the number (N) of output neurons, the method comprising the following steps:

• learning by the coding and decoding neural networks, by adapting their synaptic coefficients to examples of blocks of input data,
• encoding data blocks to be processed at the transmission end,
• transmitting encoded data via said channel,
• decoding of the encoded data received by the matched decoding neural network at the receiving end and formation of the recovered data blocks, characterized in that there is an arrangement of M coding neural networks, in competition, with each other, each identified by an index, wherein M>1, provided at the transmission end each of which effect the same type of transform and the encoded data of one or more of which are transmitted, after selection, toward paired decoding neural networks which form part of a set of several matched decoding neural networks positioned at the receiving end, each coding and decoding neural network being adapted as a result of learning to a specific class of predetermined examples of data blocks and matched with the aid of a predetermined adaptation criterion, selecting on the basis of a selection criterion, the one or more coding neural networks which effect the best coding of the data to be processed and, transmitting over the channel the data encoded by said one or more coding neural networks and their indices to render the matched decoding neural networks active.