PURPOSE: To provide a learning device for a recurrent neural network which learns the recurrent network fast with high precision by using a second-order derivative.

CONSTITUTION: An initialization part 21 initialize coupling weight Φ1 and control variables H1 and (k), and a stop condition decision part 22 decides the stop conditions of algorithm. When the stop conditions are not met, a correction direction calculation part 23 calculates a correction direction vector ΔΦk=-Hk...f(Φ), an optimum search step width calculation part 24 calculates optimum step width λk minimizing an object function f(Φk+λkΔΦk), and a coupling weight update part 25 updates the coupling weight Φk+1=Φk+λkΔΦk. An approximate matrix calculation part 26 calculates Hk+1 on the basis of the conditions and the stop condition decision part 22 decides the stop conditions.