Data compression and novelty filtering in retinotopic backpropagation networks

Three-layer networks with identical numbers of input and output units were trained using standard back-propagation to reproduce vectors of independent identically distributed random variables. Data compression was accomplished by virtue of the hidden layer´s having fewer units than the input or output. The trained nets gave a transparent response to training inputs and a translucent response when anomalies were added to elements of the training set. The reproducing vector closely resembles the unperturbed input. By subtracting the output vector from the input, a filter results, since the anomalies are dramatically enhanced in the difference vector