Circuits and method for shaping the influence field of neurons and neural networks resulting therefrom

Abstract

The improved neural network of the present invention results from the combination of a dedicated logic block with a conventional neural network based upon a mapping of the input space usually employed to classify an input data by computing the distance between said input data and prototypes memorized therein. The improved neural network is able to classify an input data, for instance, represented by a vector A even when some of its components are noisy or unknown during either the learning or the recognition phase. To that end, influence fields of various and different shapes are created for each neuron of the conventional neural network. The logic block transforms at least some of the n components (A1, . . . , An) of the input vector A into the m components (V1, . . . , Vm) of a network input vector V according to a linear or non-linear transform function F. In turn, vector V is applied as the input data to said conventional neural network. The transform function F is such that certain components of vector V are not modified, e.g. Vk=Aj, while other components are transformed as mentioned above, e.g. Vi=Fi(A1, . . . , An). In addition, one (or more) component of vector V can be used to compensate an offset that is present in the distance evaluation of vector V. Because, the logic block is placed in front of the said conventional neural network any modification thereof is avoided.