| 摘要 |
PURPOSE: To converge a parameterθof the neural network on one of real parameters by making a coefficient of a regularization item added to a loss function meet specific conditions. CONSTITUTION: For the neural network which outputs data (y) for an input of (x) by a function (f) having the parameterθ, (n) pieces of learning data xi and yi (i=1 to n) in proper combination are prepared, and a data comparing means compares the learning data yi with data (y) outputted by inputting the learning data xi to the neural network by a data input means. A parameter correcting means corrects the parameterθso that a loss function represented as an expression showing their difference and a function represented as equation II consisting of a regularization term Q(θ) become minimum. The learning device for the neural network makes the coefficient,λof the regularization term meet conditions ofλ=αn /n and 0<αn <n.
|
