Kokusai Denshin Denwa Co., Ltd.

Youtaro Yatsuzuka

Fumiaki Sugaya

Masaru Enomoto

Shigeru Iizuka

A reiterative learning procedure with training and test processes for a binary supervised neural network includes at least an error signal generator for weighting factor updating in the training process, which generates an error signal that is perturbed in polarity and amplitude in the difference derived by subtracting an output unit signal from corresponding binary teacher signal and then generates the difference as an error signal after a maximum absolute value of differences among erroneous binary output signals has become smaller than a threshold once. A training signal memory stores a set of training signals and adds test signals providing erroneous binary output signals that are transferred from a test signal memory in the test process to the set of training input signals as incremental training input signals. An affordable signal memory stores input signals with sufficiently large margins providing correct binary output signals that are transferred from the training signal memory in the training process and the test signal memory in the test process. The reiterative learning procedure, with minimum necessary training and test input signals and control of the error perturbation in the training process, can provide a binary space to obtain a desired binary output, and also realizes an extremely high generalization ability.

Learning procedure for multi-level neural network

Parallel multi-value neural networks

A learning method supervised by a binary teacher signal for a binary neural network comprises at least an error signal generator 10 for weighting factor updating, which generates an error signal for weighting factor updating having an opposite polarity to that of a difference signal between an output unit signal of the binary neural network and the binary teacher signal on an output unit whereat a binary output unit signal coincides with the binary teacher signal, and an amplitude which decreases by increase of distance from the binary teacher signal, when an absolute value of the difference signal is smaller than a threshold, generates an error signal which has the same polarity as that of the difference signal and an amplitude smaller than that of the difference signal, when the absolute value of the difference signal is larger than the threshold, or generates an error signal which has an amplitude equal to or smaller than that of the difference signal on an output unit providing a wrong binary output unit signal which is different from the binary teacher signal. Updating the weighting factors by the error signal which is optimally generated according to discrimination between the correct binary output unit signal and the wrong one, can provide a binary neural network which converges very quickly and reliably to obtain a desired binary output and also realizes a high generalization ability.

Learning method for multi-level neural network

In an echo canceller having PCM interfaces used in a PCM telephone circuit, an echo estimator (6) is coupled with a receive path through a linear code converter (13) for obtaining a linear receive input signal, and a first subtractor (7) is inserted after a linear code converter (13) in a transmit path for cancelling an echo component in a linear transmit input signal and for transmitting a first residual echo to the far-end talker. A non-linear quantization processor (24) is inserted between the first subtracter (7) and the echo estimator (6), and produces a new echo estimate containing quantization noise, based on an echo estimate which is derived from the echo estimator (6). The filter coefficients in the echo estimator (6) are updated by using the second residual echo obtained by the second subtracter (23) separately prepared from the first subtracter (7). The new echo estimate from the non-linear quantization processor (24) can drastically cancel not only the echo but also the quantization noise in the linear transmit input signal without any effect on the performance of the echo estimator (6). The echo canceller having the non-linear quantization processor (24) and subtracters separately prepared for transmitting the first residual echo to the far-end talker and for updating the echo estimator (6) can achieve a high echo return loss enhancement even for a multi-linked PCM channel.

Echo canceller using two residual echoes

In an echo canceller used in a four-wire digital telephone circuit, a main echo estimator (21) and a sub echo estimator (22) are coupled with a receive path for measuring a receive signal, and subtracters (25, 26) are inserted in a transmit path for cancelling an echo component in a transmit signal. Filter coefficients in said estimators (21, 22) are adaptively controlled by a convergence control processor (24). The main echo estimator has a small step gain for updating filter coefficients so that the response to an echo is slow, while the sub echo estimator (22) has a large step gain and quick response for an echo. An register accumulator (23) is provided so that the filter coefficients by the sub echo estimator (22) are selectively accumulated on the related filter coefficients in the main echo estimator (21) according to operation modes, ordinary mode, accumulating mode, and reset mode. In an ordinary mode, said two echo estimators (21, 22) operate independently from each other, updating the filter coefficients. In an accumulating mode, when an echo output of the second subtracter (26) is lower than that of the first subtracter (25), and the filter coefficients in the sub echo estimator (22) converge, the filter coefficients in the sub echo estimator (22) are accumulated on the related ones in the main echo estimator (21), respectively, and after reset of the coefficients in the sub echo estimator (22), the mode is then switched to the ordinary mode. When the filter coefficients in the sub echo estimator (22) diverge, or echo output of the second subtracter (26) is larger than that of the first subtracter (25), the sub echo estimator (22) is reset, and then, the mode is switched to the ordinary mode.

Growing community of inventors

Kanagawa, Japan

Youtaro Yatsuzuka