Abstract: An information processing apparatus includes a first recurrent neural network (RNN) for performing processing which corresponds to a time-series and a second RNN for processing another correlated time-series. The difference between a context set output by the first RNN and a context set output by the second RNN is computed by a subtractor, and the obtained difference is used as a prediction error. Backpropagation is performed based on the prediction error, thus determining a coefficient for each neuron of an output layer, an intermediate layer, and an input layer.

Abstract: A learning-type movement control apparatus that learns the movement of an operation control device, predicts the movement thereof, and drives it so as to automatically move. The apparatus comprises an operation control device having a predetermined portion that is displaced according to a force exerted in an arbitrary direction, outputs the amount of the displacement at least as one-dimensional position-representing information, receives a feedback signal carrying information generated by adding displacement information to the position-representing information, and drives the predetermined portion according to a direction and a displacement that are based on the feedback signal. The apparatus also includes a learning section that receives the position-representing information and performs learning of the movement of the operation control device.

Abstract: A land mark recognition apparatus presenting an enhanced learning capability with its neural net. Color image data of a land mark picked up by a television camera is quantized in the space defined by hue and chroma. Specifically, the color image data defined by hue and chroma falls within an area A1, the data is quantized as red data. Similarly, color image data falls within areas A3 and A3, the data is quantized as green and blue data, respectively. The color image data, after being quantized into a small number of data, is stored in the neural net.

Abstract: A signal processing method for efficiently searching an optimum solution in a neural network by including a term of a nonlinear resistance in an equation of motion and changing such nonlinear resistance periodically. According to the method, the range of absolute values of connection weights between units in the neural network is limited by the equation of motion, hence preventing a prolonged search time that may otherwise be caused by excessive extension of the search scope beyond the requisite. A plurality of patterns are previously embedded or stored in the neural network and, upon input of a predetermined key pattern, the nonlinear resistance is changed periodically to recall a pattern similar to the key pattern, whereby any desired pattern can be searched or retrieved with rapidity and facility out of the complicated patterns.

Abstract: According to the present invention, a method for recognition of normal and abnormal conditions can be performed with at least one neural network. First, trend data of an object system, before a recognition-step, are entered as input data to an input layer of each neural network and data of this system at the recognition-step are entered as objective output data to an output layer of the neural network. Thus, multiple sets of trend data showing at least one normal condition of this system are formed in the neural network in order to obtained learned weights and biases. Next, output data at every recognition-step are predicted by entering actual trend data as input data to the neural network, while the learned weights and biases are utilized. Then, the predicted output data are compared with actual output data at every recognition-step.

Abstract: To enable the pattern matching between a shifted input pattern and the standard pattern, a plurality of standard patterns are stored in a standard pattern associative memory network 12. A pattern shifted relative to the standard pattern is inputted to the input pattern network 11 and a restriction condition of when the input pattern is shifted relative to the standard pattern is stored in a coordinate associated network 14. In an association network 13, weights and biases are determined so that the respective units of the network 13 are activated most intensely when the input pattern and the standard pattern match correctly each other in response to the signals from the respective networks 11, 12, and 14.