Trainable neural network having short-term memory for altering input layer topology during training

• US 5,627,942 A
• Filed: 07/07/1992
• Issued: 05/06/1997
• Est. Priority Date: 12/22/1989
• Status: Expired due to Term

• Current Assignee: British Telecommunications (BT Group Plc) (BT Group PLC)
• Original Assignee: British Telecommunications PLC (BT Group PLC)
• Inventors: Wyard, Peter J., Nightingale, Charles
• Primary Examiner(s): Davis, George B.
• Application Number:
  US07/867,698
• Time in Patent Office: 1,764 Days
• Field of Search: 395/23, 395/24, 395/20
• US Class Current: 706/25
• CPC Class Codes: G06F 18/00 Pattern recognition
• CPC Class Codes: G06N 3/082 modifying the architecture, e.g. adding, deleting or silencing nodes or connections

Abstract

A neutral net in which new nodes and connections are created in both input and intermediate layers during training, which is by punishment, reward and teaching. This can use a small increase in memory requirement to preclude the necessity for long training times applicable problems in speech and natural language processing, video recognition and simple logic functions.

Claims

1. A trainable artificial neural network comprising:

input layer means for inputting sets of data items and generating node outputs, each node output corresponding to an occurrence anywhere in a set of a respective predetermined data item;

output layer means for responding to the generation of said node outputs by producing at least one net output, each net output depending upon a plurality of such node outputs weighted by corresponding weight values;

means for storing given data items in said set, each given data item stored not corresponding to any said respective predetermined data item; and

means for causing said input layer means to thereafter generate node outputs corresponding to occurrences in a said set of respective said given data items.

2. A network according to claim 1, in which each said set of data items is in the form of a sequence of the corresponding data item and said given data items are predetermined subsequences of said predetermined data items in said sequence.

3. A network according to claim 2 in which said predetermined subsequences comprise predetermined data items occurring contiguously in said sequence.

4. A network according to claim 2 in which said predetermined subsequences comprise fewer than a predetermined number of said predetermined data items.

5. A network according to claim 2 in which said predetermined subsequences are a subset of subsequences of said predetermined data items occurring in said sequence, the probability of a given subsequence being included in the subset decreasing with increasing subsequence length.

6. A network according to claim 1 further comprising means for inputting sequences of data items and generating said sets of data items, each generated set of data items corresponding to an inputted said sequence and comprising representations of respective data items in the corresponding sequence, each representation being indicative of both the identity of the data item it represents and the position of the data item in the sequence.

7. A network according to claim 1 further comprising control means for inhibiting said means for causing in dependence upon variable criteria.

8. A network according to claim 7 in which one of said variable criteria corresponds to the number of node or intermediate outputs.

9. A network according to claim 8, in which a criterion corresponds to the number of intermediate outputs to which a given output means output is responsive, whereby each given output means output is constrained not to receive more than a predetermined number of intermediate outputs.

10. A network according to claim 8 in which a criterion corresponds to the number of node outputs already available so as to constrain the input layer means not to produce more than a predetermined number of such node outputs.

11. A network according to claim 7 in which one such criterion corresponds to a recent historical performance of the network, such that a good historical performance will tend to inhibit said means for causing.

12. A network according to claim 1 including means for causing said input layer means to cease to generate a node output corresponding to an occurrence anywhere in a said set of a data item which has, over a relatively long interval, not been present in inputted sets of data items.

13. A network according to claim 1 in which the means for storing comprises means for temporarily storing said data and in which the input layer means comprises means for temporarily storing any said respective predetermined data item occurring in said set of data items, said network comprising a digital computer operating under stored program control.

14. A network according to claim 1 wherein each layer means comprises a plurality of nodes each comprising at least one input and an output, and means for producing a node output signal in dependence upon each input.

15. A method of training an artificial neural network having input layer means for inputting sets of data items and generating node outputs, each node output corresponding to an occurrence anywhere in a set of a respective predetermined date item, and output layer means for responding to the generation of said node outputs by producing at least one net output, each net output depending upon a plurality of such node outputs weighted by corresponding weight values the method comprising the steps of:

inputting training sets of data to the input layer means;

storing given data items in a training set, each stored given data item not corresponding to any said respective predetermined data item;

detecting said net output, and, in the event that a predetermined criterion of success is not met, causing said input layer means to thereafter generate node outputs corresponding to occurrences in said sets of data items of at least one said given data item.

16. A method of training a neural network having network inputs and network outputs defined by a plurality of weight values, said method comprising the steps of:

inputting training data and adjusting said weight values as a function of network outputs, an amount by which a given weight value is adjusted depending upon said given weight value such that the amount is less for both high weight values and low weight values than for weight values intermediate said high weight values and low weight values, whereby after frequent adjustments in a first direction, adjustment in opposite second direction has little effect and, after frequent adjustments in said opposite second direction adjustment in said first direction has little effect.

17. A method according to claim 16, wherein each weight W_old is adjusted by a increment Δ

given by the formula;

##EQU3## where μ

is a constant and δ

=±

1.