Neural network having an optimized transfer function for each neuron

• US 5,280,564 A
• Filed: 02/18/1992
• Issued: 01/18/1994
• Est. Priority Date: 02/20/1991
• Status: Expired due to Fees

• Current Assignee: Honda Giken Kogyo Kabushiki Kaisha (Honda Motor Company)
• Original Assignee: Honda Giken Kogyo Kabushiki Kaisha (Honda Motor Company)
• Inventors: Shiomi, Kazuyuki, Watanabe, Sei
• Primary Examiner(s): Fleming, Michael R.
• Assistant Examiner(s): Hafiz, Tariq R.
• Application Number:
  US07/836,593
• Time in Patent Office: 700 Days
• Field of Search: 395/23, 395/24, 395/27
• US Class Current: 706/25
• CPC Class Codes: G06N 3/048 Activation functions
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

The characteristic data for determining the characteristics of the transfer functions (for example, sigmoid functions) of the neurons of the hidden layer and the output layer (the gradients of the sigmoid functions) of a neural network are learned and corrected in a manner similar to the correction of weighting data and threshold values. Since at least one characteristic data which determines the characteristics of the transfer function of each neuron is learned, the transfer function characteristics can be different for different neurons in the network independently of the problem and/or the number of neurons, and be optimum. Accordingly, a learning with high precision can be performed in a short time.

Claims

1. A neural network comprising a plurality of neurons each performing a signal processing corresponding to a neural element, said neurons being hierarchically connected in the order of at least one input layer, at least one hidden layer and an output layer, and said input layer, hidden layer and output layer each consisting of at least one neuron, individual neurons of said hidden layer and output layer being operative to correct the data weighted by multiplying the outputs of the preceding input layer or hidden layer by predetermined weighting data with a predetermined threshold value, respectively, and to substitute the data after the correction in a predetermined transfer function to calculate output data, said neural network including:

error data calculation means operative to use a predetermined cost function to calculate error data from the output data of the output layer,

weighting data correction means operative to partially differentiate said cost function with each of a plurality of weighting variables to obtain partial differentiated coefficients, and to correct said weighting data according to said partial differential coefficients.

threshold value correction means operative to partially differentiate said cost function with each of a plurality of threshold variables to obtain partial differentiated coefficients, and to correct each said threshold value according to said partial differential coefficients, and

characteristic data correction means for partially differentiating said cost function, to obtain partial differentiated coefficients, with at least one characteristic variable that determines the characteristics of the respective transfer functions of said hidden layer and said output layer, and operative to correct said characteristic data according to said partial differential coefficients.

2. A neural network as set forth in claim 1 wherein said characteristic data correction means corrects said characteristic data using the partial differential coefficients obtained by partially differentiating said cost function with said at least one characteristic variable, and the previous values of said partial differential coefficients.

3. A neural network as set forth in claim 2 wherein said cost function is an error function, and said error data calculation means compares the data output from said output layer with training data corresponding to the input data supplied to said input layer and applies the comparison result to said error function, thereby to calculate error data.

4. A neural network as set forth in claim 2 wherein said weighting data correction means uses the partially differentiated value of said cost function by said weighting variables and the previous value of said partially differentiated value to correct said weighting data.

5. A neural network as set forth in claim 2 wherein said threshold value correction means uses the partially differentiated variables of said cost function by said threshold value and the previous value of said partially differentiated value to correct said threshold value.

6. A neural network as set forth in claim 1 wherein said transfer function is a sigmoid function and said characteristic variable is the gradient of the sigmoid function.

7. A neural network as set forth in claim 1 wherein said transfer function is a power function (ax^b +c) and said characteristic variable is at least one constant of said power function.

8. A neural network as set forth in claim 1 wherein said transfer function is a trigonometric function {a Sin(cx)+b Cos(dx)} and said characteristic variable is at least one constant of said trigonometric function.

9. A neural network as set forth in claim 1 wherein said cost function is an error function, and said error data calculation means compares the data output from said output layer with training data corresponding to the input data supplied to said input layer and applies the comparison result to said error function, thereby to calculate error data.

10. A neural network as set forth in claim 1 wherein said weighting data correction means uses the partially differentiated value of said cost function by said weighting variables and the previous value of said partially differentiated value to correct said weighting data.

11. A neural network as set forth in claim 1 wherein said threshold value correction means uses the partially differentiated value of said cost function by said threshold variables and the previous value of said partially differentiated value to correct said threshold value.