Valley searching method for recurrent neural networks

The authors present a new learning algorithm called the VSM (valley searching method) for the supervised learning of RNNs (recurrent neural networks). H. Akaike had originally proposed to accelerate a search for a minimum of the quadratic function with a positive definite symmetric matrix. It is shown that VSM is very effective for searching for a minimum in the shape of the curved narrow valley peculiar to the RNN learning surface where learning is executed