Data processing using neural networks having conversion tables in an intermediate layer

• US 6,243,490 B1
• Filed: 11/13/1997
• Issued: 06/05/2001
• Est. Priority Date: 06/14/1990
• Status: Expired due to Fees

• Current Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Original Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Inventors: Mita, Yoshinobu
• Primary Examiner(s): CHANG, JON CARLTON
• Application Number:
  US08/969,832
• Time in Patent Office: 1,300 Days
• Field of Search: 382/156, 382/157, 382/158, 706/15, 706/41, 706/42, 706/43, 706/26, 364/757, 364/760.03, 364/148.03
• US Class Current: 382/158
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

In a neural network which includes one input layer, one or more intermediate layers and one output layer, neural elements in the input layer and neural elements in the intermediate layer are divided into groups. Arithmetic operations representing the coupling between the neural elements of the input layer and the neural elements of the intermediate layer are put into table form.

Claims

1. A method of concretely constructing a neural network in which coupling coefficients have already been decided by learning, comprising the steps of:

determining, with respect to each coupling coefficient of one or more neural elements contained in said neural network, a sum of a plurality of terms of powers of 2 so that the number of the plurality of terms is minimized under the condition that a difference between the sum and the coefficient is not greater than a predetermined allowable error; and

replacing multiplication between a coupling coefficient and input data inputted to the neural element thereof by processing for shifting said input data by each exponent for every one of a plurality of exponents of the obtained powers of 2, and addition processing for adding the results of shifting.

2. The method according to claim 1, wherein said network has an input layer, one or more intermediate layers, and an output layer, and said multiplication is performed for the sake of coefficients between said input layer and said intermediate layer.

3. The method according to claim 1, wherein said processing for shifting said input data by each exponent for every one of a plurality of exponents of the obtained powers of 2 is executed by regarding the inputted data in said addition processing as data of value after the shifting.

4. A neural network constructed by the method according to claim 1.

5. A method of concretely constructing a neural network in which coupling coefficients have already been decided by learning, comprising the steps of:

approximating each coupling coefficient of a plurality of neural elements contained in said neural network by a term of a power of 2;

replacing multiplication between each coupling coefficient of the plurality of neural elements and input data inputted to this neural element by processing for shifting the input data by an exponent of the obtained power of 2; and

adding all the shifted data by one or more adders by integrating into a single addend a plurality of shifted data which do not share the same digit positions among the shifted data arising from different neural elements, and inputting the integrated addend to one input terminal of one of the adders.

6. The method according to claim 5, wherein the coupling coefficients of remaining neural elements are relearned after said multiplication is replaced by shift processing.

7. A neural network constructed by the method according to claim 5.

8. A method of concretely constructing a neural network in which coupling coefficients have already been decided by learning, comprising the steps of:

approximating each coupling coefficient of a plurality of neural elements contained in said neural network by a sum of a plurality of terms of powers of 2;

replacing multiplication between each coupling coefficient of the plurality of neural elements and input data inputted to this neural element by processing for shifting said input data by an exponent for every one of a plurality of exponents of the obtained powers of 2, and addition processing for adding the results of shifting; and

in an adder having two input terminals used in this addition processing, integrating into a single addend a plurality of shifted data which do not share the same digit positions among the shifted data arising from different neural elements, and inputting the integrated addend to one of the input terminals of the adder.

9. The method according to claim 8, further comprising the step of inputting, to respective input terminals of the adder, two items of shifted data at the same digit positions in shifted data arising from two or more different neural elements.

10. The method according to claim 8, wherein the coupling coefficients of remaining neural elements are relearned after said multiplication is replaced by shift processing and addition processing.

11. The method according to claim 8, wherein when the coupling coefficient of each of the plurality of neural elements is approximated by a sum of a plurality of terms of powers of 2, the plurality of terms is determined so that the number of terms is minimized under the condition that a difference between the sum and the coefficient is not greater than a predetermined allowable error.

12. A neural network constructed by the method according to claim 8.