Provides are a neural network learning method, a neural network generation method, a learned device, a portable terminal device, a learning processing device, and a computer program that do not require use of a backward propagation method. The present invention involves: setting neurons in a neural network as random variables each capable of taking two values; expressing connection weights between the respective neurons by using a plurality of synapses each of which has undergone multiplication by a required connection coefficient; setting the plurality of synapses as random variables each capable of taking two values; giving initial data to neurons in an intermediate layer; repeating a process in which sampling based on the Markov chain Monte Carlo method is performed on a conditional probability distribution, under a condition that the random variables of neurons in an input layer and an output layer represent training data values, and the status values of the synapses and the neurons in the intermediate layer are updated; and calculating, on the basis of the updated status values of the respective synapses, the connection weights between the respective neurons.
展开▼
