Machine-based learning system

• US 5,212,821 A
• Filed: 03/29/1991
• Issued: 05/18/1993
• Est. Priority Date: 03/29/1991
• Status: Expired due to Term

• Current Assignee: American Telephone & Telegraph Company (AT&T, Inc.)
• Original Assignee: AT&T, Inc.
• Inventors: Levinson, Stephen E., Gorin, Allen L.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/677,043
• Time in Patent Office: 781 Days
• Field of Search: 395/2, 395/22, 395/23, 381/43, 382/14, 382/15
• US Class Current: 706/20
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A machine-based learning system receives inputs and proposes outputs using a network of connection weights to relate the inputs to the outputs. When a user affirms a proposed output, the system adjusts the connection weights to strengthen the relationship between elements of the inputs and that output. Learning occurs in the course of successive iterations as the connection weights adapt to relationships between semantically-significant input elements and related outputs. The system can be used to acquire language from inputs provided in the form of text or speech such that the connection weights adapt to semantically-significant words without need for defining word meanings. Such language acquisition contrasts with previous systems in which words are predefined.

Claims

1. The method of conditioning a machine for selecting one of a plurality of outputs in response to applied inputs, each input containing one or more elements, comprising the steps of:

establishing a recognition node for each distinct element in said inputs, said recognition node having activation states indicating the presence or absence of said element in an input;

defining weighting relationships between each said recognition node and each said output;

determining, in response to each input, a total activation for each said output from said weighting relationships and the activation states of said recognition nodes;

selecting, in response to said determining step, the output having the most favorable total activation;

proposing the output selected in said selecting step;

adjusting said weighting relationships in response to an input received subsequent to said proposing step in accordance with whether or not said subsequent input indicates that said proposed output is correct;

whereby, as successive inputs are received, said weighting relationships become adapted to represent associations between semantically significant ones of said elements and said outputs.

2. The method of claim 1 wherein said activation states of said recognition nodes are expressed as probabilities and said determining step further comprises calculating said total activations as sums of products of said weighting relationships and said probabilities.

3. The method of claim 1 wherein said adjusting step further comprises calculating said weighting relationships to be the mutual information I(c_k, v_m) between said recognition nodes v_m and said actions c_k and wherein

4. The method of claim 1 wherein said inputs further comprise text and said elements further comprise words.

5. The method of claim 4 wherein said text further comprises spoken utterances and said words further comprise spoken words.

6. The method of claim 4 wherein said establishing step further comprises establishing recognition nodes for adjacent words in said text.

7. The method of claim 1 which includes the additional step of defining at least one of said elements to signify a correct proposed output.

8. The method of claim 1 which includes the additional step of defining the expiration of a waiting period after said proposing step to signify a correct proposed output.

9. The method of claim 1 wherein said proposing step further comprises displaying text.

10. The method of claim 1 wherein said proposing step further comprises reciting speech.

11. The method of claim 1 wherein said establishing step further comprises establishing recognition nodes for groups of said elements.

12. The method of claim 1 which includes the additional step of defining at least one of said elements to signify an incorrect proposed output.

13. The method of claim 1 wherein said establishing step further comprises establishing recognition nodes that respond to the activation states of at least two other recognition nodes.

14. The method of claim 1 including the additional step of, before the proposing step, inhibiting actions proposed in response to previous inputs.

15. The method of claim 1 including the additional step of initiating said proposed output if said subsequent input indicates that said proposed output is correct.

16. A machine-based system for selecting one of a plurality of outputs in response to inputs, each said input containing one or more elements, and for learning associations between said elements and said actions, which comprises:

means for receiving said inputs;

means responsive to said inputs for establishing a recognition node for each distinct element in said inputs, said recognition node having activation states indicating the presence or absence of said element in an input;

means for defining weighting relationships between each said recognition node and each said output;

means for determining, in response to receiving each input, a total activation for each said output from said weighting relationships and the activation states of said recognition nodes

means responsive to said determining means for selecting the output having the most favorable total activation;

means responsive to said selecting means for proposing the selected action;

means for adjusting said weighting relationships in response to an input received subsequent to proposing said output in accordance with whether or not said subsequent input indicates that said proposed output is correct;

whereby, as successive inputs are received, said weighting relationships become adapted to represent associations between semantically significant ones of said elements and said outputs.

17. The system of claim 16 wherein the means for establishing also comprises means for establishing recognition nodes that respond to the activation states of at least two other recognition nodes.

18. The system of claim 16 which further comprises means for initiating a proposed output if said subsequent input indicates that such output is correct.

19. The system of claim 16 in which said means for establishing, means for defining, means for determining, means for selecting and means for adjusting collectively further comprise a programmed digital computer.