We describe an approach used for the conception of neural real-time pattern recognition systems which are interruptible, that is, able to give answers before the computing completion. This method is based on dynamic coding of information in the neural network: the proposed neuron model performs a time integration of its inputs and emits binary spikes trains. We have designed a dynamic shared-weights multi-layer neural classifier with recognition rate close to a more classical pattern recognition netowrk, suitable for real-time applications with early hypothesis production. Because of its low computing complexity, this model seems to be well suited to hardware implementations.
展开▼
