Learning machine with a hierarchial structure without mutual connections within levels thereof

A learning machine having an output signal calculation means comprising plural multiple-input single-output signal processors network connected in a hierarchical structure without mutual connections within levels in such a manner that signals propagate only to a higher level, and a weight coefficient renewing means which renews values of weight coefficients of the output signal calculation means based on output signals obtained by the output signal calculation means;

• each multiple-input single-output signal processor comprising a memory to store plural weight coefficients, an input section for inputting plural data, multiplying means to weight the input data from the input section by the weight coefficients stored in the memory, an adding means to add the plural data weighted by the multiplying means, and a threshold value processor to limit an output of the adding means to a value within a predetermined range;
• and
• the weight coefficient renewing means comprising a teacher signal generator to produce as a teacher signal a desirable value for the output signal of the output signal calculation means, an error signal calculation means to obtain an error between the output signal and the teacher signal, a steepest descent direction determination means to obtain a direction of steepest descent from an output of the error signal calculation means;
• the weight coefficient renewing means further comprising an amount of weight change calculation means which calculates an amount of change in the weight coefficient from the direction of steepest descent, a linear search means which outputs amounts of weight change to plural learning rates for the direction of steepest descent, a parabolic line approximation means which determines an amount of weight change by parabolic line approximation method from the plural weight change amounts output from the linear search means, a learning progress evaluation means evaluates progress of learning from the error output by the error signal calculation means, a threshold value control means which gradually reduces a threshold value as learning progresses according to an evaluation result of the learning progress evaluation means, a first error signal evaluation means which evaluates whether the output of the error signal calculation means is less than the threshold value of the threshold value control means, a first amount of weight change control means which sets the amount of weight coefficient change to 0 when the error is determined to be less than the threshold value by the first error signal evaluation means, an amount of weight change selection means which selects either an output of the amount of weight change calculation means or an output of the parabolic line approximation means according to the level of the threshold value, and a weight modification means which changes the value of the weight coefficient stored in the memory according to an amount of weight change selected by the amount of weight change selection means.

A learning machine having an output signal calculation means comprising plural multiple-input single-output signal processors network connected in a hierarchical structure without mutual connections within levels in such a manner that signals propagate only to a higher level, and a weight coefficient renewing means which renews values of weight coefficients of the output signal calculation means based on output signals obtained by the output signal calculating means;

• each multiple-input single-output signal processor comprising a memory to store plural weight coefficients, an input section for inputting plural data, multiplying means to weight the input data from the input section by the weight coefficients stored in the memory, an adding means to add the plural data weighted by the multiplying means, and a threshold value processor to limit an output of the adding means to a value within a predetermined range;
• and
• the weight coefficient renewing means comprising a teacher signal generator to produce as a teacher signal a desirable value for the output signal of the output signal calculation means, an error signal calculation means to obtain an error between the output signal and the teacher signal, a steepest descent direction determination means to obtain a direction of steepest descent from an output of the error signal calculating means;
• the weight coefficient renewing means further comprising an amount of weight change calculation means which calculates an amount of change in the weight coefficient from the direction of steepest descent, a conjugate gradient direction determination means to obtain a direction of the conjugate gradient from the direction of steepest descent, a linear search means which outputs amounts of weight change to plural learning rates for the direction of the conjugate gradient, a parabolic line approximation means which determines an amount of weight change by parabolic line approximation method from the plural weight change amounts output from the linear search means, a learning progress evaluation means which evaluates progress of learning from the error output by the error signal calculation means, a threshold value control means which gradually reduces a threshold value as learning progresses according to an evaluation result of the learning progress evaluation means, a first error signal evaluation means which evaluates whether the output of the error signal calculation means is less than the threshold value, a first amount of weight change control means which sets the amount of weight coefficient change to 0 when the error is determined to be less than the threshold value by the first error signal evaluation means, an amount of weight change selection means which selects either an output of an amount of weight change calculation means or an output of the parabolic line approximation means according to the level of the threshold value, and a weight modification means which changes the value of the weight coefficient stored in the memory according to an amount of weight change selected by the amount of weight change selection means.

A learning machine having an output signal calculation means comprising plural multiple-input single-output signal processors network connected in a hierarchical structure without mutual connections within levels in such a manner that signals propagate only to a higher level, and a weight coefficient renewing means which renews values of weight coefficients of the output signal calculation means based on output signals obtained by the output signal calculation means;

• each multiple-input single-output signal processor comprising a memory to store plural weight coefficients, an input section for inputting plural data, multiplying means to weight the input data from the input section by the weight coefficients stored in the memory, an adding means to add the plural data weighted by the multiplying means, and a threshold value processor to limit an output of the adding means to a value within a predetermined range;
• and
• the weight coefficient renewing means comprising a teacher signal generator to produce as a teacher signal a desirable value for the output signal of the output signal calculation means, an error signal calculation means to obtain an error between the output signal and the teacher signal, a steepest descent direction determination means to obtain a direction of steepest descent from an output of the error signal calculating means;
• the weight coefficient renewing means further comprising a conjugate gradient direction determination means to obtain a direction of a conjugate gradient from the direction of steepest descent, a learning progress evaluation means which evaluates progress of learning from the error output by the error signal calculation means, a threshold value control means which gradually reduces a threshold value as learning progresses according to an evaluation result of the learning progress evaluation means, a first error signal evaluation means which evaluates whether the output of the error signal calculation means is less than the threshold value, a skip evaluation means which counts the outputs of the first error signal evaluation means and outputs a skip signal when all output signals are less than the threshold value, a weight change direction selection means which selects either the direction of the conjugate gradient or the direction of the steepest descent as the direction of weight change according to the level of the threshold value, an amount of weight change calculation means which calculates an amount of change in the weight coefficient from the selected direction of weight change, a weight modification means which changes the value of the weight coefficient stored in the memory according to the output of the amount of weight change calculation means, and a weight change control means which causes the weight modification means to skip a weight change operation according to the skip signal.

A learning machine having an output signal calculation means comprising plural multiple-input single-output signal processors network connected in a hierarchical structure without mutual connections within levels in such a manner that signals propagate only to a higher level, and a weight coefficient renewing means which renews values of weight coefficients of the output signal calculation means based on output signals obtained by the output signal calculation means;

• each multiple-input single-output signal processor comprising a memory to store plural weight coefficients, an input section for inputting plural data, multiplying means to weight the input data from the input section by the weight coefficients stored in the memory, an adding means to add the plural data weighted by the multiplying means, and a threshold value processor to limit an output of the adding means to a value within a predetermined range;
• and
• the weight coefficient renewing means comprising a teacher signal generator to produce as a teacher signal a desirable value for the output signal of the output signal calculation means, an error signal calculation means to obtain an error between the output signal and the teacher signal, a steepest descent direction determination means to obtain a direction of steepest descent from an output of the error signal calculation means;
• the weight coefficient renewing means further comprising an amount of weight change calculation means which calculates an amount of change in the weight coefficient from the direction of steepest descent, a linear search means which outputs amounts of weight change to plural learning rates for the direction of steepest descent, a parabolic line approximation means which determines an amount of weight change by parabolic line approximation method from the plural weight change amounts output from the linear search means, a learning progress evaluation means which evaluates progress of learning from the error output by the error signal calculation means, a threshold value control means which gradually reduces the threshold value as learning progresses according to an evaluation result of the learning progress evaluation means, a first error signal evaluation means which evaluates whether the output of the error signal calculation means is less than the threshold value, a skip evaluation means which counts the outputs of the first error signal evaluation means and outputs a skip signal when all output signals are less than the threshold value, an amount of weight change selection means which selects either an output of the amount of weight change calculation means or an output of the parabolic line approximation means according to the level of the threshold value, a weight modification means which changes the value of the weight coefficient stored in the memory according to an amount of weight change selected by the amount of weight change selection means, and a weight change control means which causes the weight modification means to skip a weight change operation according to the skip signal.

A learning machine having an output signal calculation means comprising plural multiple-input single-output signal processors network connected in a hierarchical structure without mutual connections within levels in such a manner that signals propagate only to a higher level, and a weight coefficient renewing means which renews values of weight coefficients of the output signal calculation means based on output signals obtained by the output signal calculation means;

• each multiple-input single-output signal processor comprising a memory to store plural weight coefficients, an input section for inputting plural data, multiplying means to weight the input data from the input section by the weight coefficients stored in the memory, an adding means to add the plural data weighted by the multiplying means, and a threshold value processor to limit an output of the adding means to a value within a predetermined range;
• and
• the weight coefficient renewing means comprising a teacher signal generator to produce as a teacher signal a desirable value for the output signal of the output signal calculation means, an error signal calculation means to obtain an error between the output signal and the teacher signal, a steepest descent direction determination means to obtain a direction of steepest descent from an output of the error signal calculation means;
• the weight coefficient renewing means further comprising an amount of weight change calculation means which calculates an amount . of change in the weight coefficient from the direction of steepest descent, a conjugate gradient direction determination means to obtain a direction of the conjugate gradient from the direction of steepest descent, a linear search means which outputs amounts of weight change to plural learning rates for the direction of the conjugate gradient, a parabolic line approximation means which determines an amount of weight change by parabolic line approximation method from the plural weight change amounts output from the linear search means, a learning progress evaluation means which evaluates progress of learning from the error output by the error signal calculation means, a threshold value control means which gradually reduces a threshold value as learning progresses according to an evaluation result of the learning progress evaluation means, a first error signal evaluation means which evaluates whether the output of the error signal calculation means is less than the threshold value, a skip evaluation means which counts outputs of the first error signal evaluation means and outputs a skip signal when all output signals are less than the threshold value, an amount of weight change selection means which selects either an output of the amount of weight change calculation means or an output of the parabolic line approximation means according to the level of the threshold value, a weight modification means which changes the value of the weight coefficient stored in the memory according to an amount of weight change selected by the amount of weight change selection means, and a weight change control means which causes the weight modification means to skip a weight change operation according to the skip signal.