To provide a semiconductor device which is functioned as a neural network device having reliability in learning and having an arithmetic speed good for a practical use, and a learning method therefor.
In the semiconductor device and the learning method to be functioned as the neural network device equipped with a neuro element having an integrator as a synapse to change a load in accordance with an inputted load voltage for processing analog data, the method has a step A for inputting prescribed input data to the neuro element and calculating an error between the target value of an output of the neuro element to the input data and a real output, a step B for calculating the change amount of the error by changing the load of the integrator later, and a step C for changing the load of the integrator on the basis of the change amount of the error. In the steps B and C, after a reset voltage Vres for aligning loads into approximately fixed value (Pres) is inputted to the integrator as a load voltage, a load voltage VD2 corresponding to a load to be changed is inputted, so that a load (P) is changed.
TOYODA KENJI
OTSUKA TAKASHI
MORITA KIYOYUKI
Takaishi Satoru
Yasu Furukawa
Toshio Nishitani
Keiji width