System for spatial and temporal pattern learning and recognition

A method, based on a neural-network simulation, for recognizing input signals sequentially measured by sensory receptors from a pattern of interest at predetermined time steps, comprising the steps of:

• storing a plurality of sensory input vectors, each representing the input signals measured at one of said predetermined time steps, in a sensory window memory matrix;
• establishing a multiplicity of synaptic pairs for receiving signals from the sensory window memory matrix in a neural processor simulator by selecting pairs of individual synapses in said neural processor simulator and by assigning random connections between each synapse in said pairs of individual synapses and each individual element of said sensory window memory matrix;
• assigning a chosen number of synaptic pairs to each of a multiplicity of neural levels in each of a multiplicity of neurons in said neural processor simulator;
• selecting a neuron for learning a pattern of interest during a learning mode of operation of said neural processor simulator;
• sequentially computing ratios, at predetermined recording intervals while signals from the pattern of interest are being measured by said sensory receptors during the learning mode of operation, of signals received from said sensory window memory matrix by all synaptic pairs in one neural level of said selected neuron, storing the ratios so calculated in a synaptic ratio matrix, and repeating this step for all levels in the neuron;
• computing, at predetermined time intervals during a recognizing mode of operation of said neural processor simulator, an instantaneous activation of each synaptic pair in the neural processor simulator, said instantaneous activation being the difference between a corresponding value stored in said ratio matrix and a ratio calculated from signals currently received from said sensory window memory matrix;
• computing, at each of said predetermined time intervals, a level activation for each level of each neuron in the neural processor simulator, said level activation being a sum of all synaptic-pair instantaneous activations in the level;
• computing an integrated level activation for each level of each neuron in the neural processor simulator over a selected integration period, said integrated level activation being a sum of all level activation values over the selected integrated period;
• computing a weighted level activation for each level by performing spatiotemporal summation of all integrated level activations at each selected integration period;
• computing a preliminary neural activation for each neuron for the selected integration period by selecting a maximum value of all of the individual weighted level activations for the neuron;
• computing an output neural activation from each neuron for the selected integration period by calculating an arithmetic average of all preliminary neural activation values over all neurons, by subtracting that average from each of individual preliminary neural activation values, and by setting negative differences equal to zero;
• determining a neural processor simulator output by comparing all output neural activations so calculated to an output threshhold value and identifying a neuron, if any, with an output neural activation greater than said output threshhold value;
• and
• identifying a learned pattern associated with said neuron, if any, with an output neural activation greater than said output threshhold value.