Student neural network

• US 20030233335A1
• Filed: 06/17/2002
• Published: 12/18/2003
• Est. Priority Date: 06/17/2002
• Status: Active Grant

• Current Assignee: The Boeing Co.
• Original Assignee: The Boeing Co.
• Inventors: Mims, AJ
• Granted Patent: US 7,062,476 B2
• Time in Patent Office: Days
• US Class Current: 706/26
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A student neural network that is capable of receiving a series of tutoring inputs from one or more teacher networks to generate a student network output that is similar to the output of the one or more teacher networks. The tutoring inputs are repeatedly processed by the student until, using a suitable method such as back propagation of errors, the outputs of the student approximate the outputs of the teachers within a predefined range. Once the desired outputs are obtained, the weights of the student network are set. Using this weight set the student is now capable of solving all of the problems of the teacher networks without the need for adjustment of its internal weights. If the user desires to use the student to solve a different series of problems, the user only needs to retrain the student by supplying a different series of tutoring inputs.

Claims

1. A neural network system comprising:

a teacher neural network, said teacher neural network capable of receiving and processing a problem set so as to produce a series of tutoring inputs, said tutoring inputs including an output solution to said problem set; and

a student neural network trained using said tutoring inputs to produce a student output that closely approximates said output solution.

2. The neural network system of claim 1, wherein said teacher neural network comprises at least one teacher input node and at least one teacher output node, said teacher input node and said teacher output node interconnected in accordance with teacher weights set to permit said teacher neural network to arrive at said output solution.

3. The neural network system of claim 2, further comprising at least one hidden node.

4. The neural network system of claim 2, wherein said tutoring inputs comprise said teacher network weight settings, said output solution, and said problem set input.

5. The neural network system of claim 4, wherein said student is trained by repeatedly processing said tutoring inputs until values of said student weights that produce said student output approximating said output solution are obtained, said student weights being fixed at said values.

6. The neural network system of claim 5, wherein said training is accomplished through back propagation of errors.

7. The neural network system of claim 5, wherein said student processes an input using said fixed student weights to produce said student output.

8. The neural network system of claim 1, wherein said tutoring inputs are introduced to said student neural network using a suitable conducting device external to said student network.

9. The neural network system of claim 1, wherein said trained student network is retrained by repeatedly processing a different set of tutoring inputs derived from a different teacher network.

10. The neural network system of claim 1, wherein said teacher network is comprised of more than one teacher network.

11. The neural network system of claim 10, wherein said student network is trained using tutoring inputs from said more than one teacher network to produce a student output that closely approximates an output solution produced by any one of said more than one teacher network.

12. A method of using a student neural network to produce student outputs that closely approximate an output solution produced by a teacher neural network comprising:

providing a teacher network with teacher weights set to produce an output solution in response to a specific problem set input;

transferring said teacher weights, said problem set input, and said output solution to said student network having student weights; and

training said student network using said teacher weights, said problem set input, and said output solution so as to arrive at a setting for said student weights that causes said student network to produce a student output that closely approximates said output solution.

13. The method of claim 12, wherein said teacher weights, said problem set input, and said output solution are transferred to said student network by way of a conducting device external to said student network.

14. The method of claim 12, wherein said training step is accomplished by back propagation of errors.

15. The method of claim 12, wherein said student network may be retrained through the introduction of a second set of teacher weights, a second problem set input, and a second output solution all produced by a second teacher network.

16. The method of claim 12, wherein said teacher network is comprised of more than one teacher network.

17. The method of claim 16, wherein said student network is trained using an output solution, a problem set input, and teacher weights from said more than one teacher network to produce a student output that closely approximates an output solution produced by any one of said more than one teacher network.