Learning of associative memory in form of neural network suitable for connectionist model

• US 5,524,177 A
• Filed: 07/02/1993
• Issued: 06/04/1996
• Est. Priority Date: 07/03/1992
• Status: Expired due to Term

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: Suzuoka, Takashi
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US08/085,050
• Time in Patent Office: 1,068 Days
• Field of Search: 395/22, 395/23, 395/24, 395/21
• US Class Current: 706/18
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

The learning of an associative memory suitable for the connectionist model which can deal with the patterns having the non-random frequencies of the appearances or the non-random correlations. In this invention, the learning of the associative memory in a form of a neural network, in which a plurality of nodes having activation values are connected by a plurality of links having link weight values, is achieved by entering a plurality of learning patterns sequentially, where each learning pattern has a plurality of elements in correspondence with the nodes, calculating an energy E of the entered learning pattern, determining a learning amount δ for the entered learning pattern according to a difference between the calculated energy E and a predetermined reference energy level Eth, and updating the link weight values of the links according to the entered learning pattern and the determined learning amount δ.

Claims

1. An apparatus, comprising:

an associative memory in a form of a neural network, including;

a plurality of nodes having activation values; and

a plurality of links, connected with the nodes, having link weight values;

pattern entering means for sequentially entering a plurality of learning patterns into the neural network, each learning pattern having a plurality of elements in correspondence with the nodes;

energy calculation means for calculating an energy E of said each learning pattern entered by the pattern entering means;

learning amount determination means for determining a learning amount δ

for said each learning pattern entered by the pattern entering means, according to a difference between the energy E calculated by the energy calculation means and a predetermined reference energy level Eth, using the following equation;

space="preserve" listing-type="equation">δ

=g(E-Eth)

where g is an upper and lower bounded monotonically increasing function with g(0)=0; and

link weight value updating means for updating the link weight values of the links according to said each learning pattern entered by the pattern entering means and the learning amount δ

determined by the learning amount determination means.

2. The apparatus of claim 1, wherein the energy calculation means calculates the energy E of said each learning pattern according to values of the elements of said each learning pattern, the link weight values previously updated by the link weight value updating means, and activation probabilities for the values of the elements in said each learning pattern to take activated values.

3. The apparatus of claim 2, wherein the energy calculation means calculates the energy E of said each learning pattern according to the following equation:

##EQU11## where k is a non-zero constant coefficient having a sign identical to that of the predetermined reference energy level Eth, Wji is a link weight value for a link between a j-th node and an i-th node, V_j is a value of a j-th element in said each learning pattern, V_i is a value of an i-th element in said each learning pattern, a_j is an activation probability for the value V_j of the j-th element in said each learning pattern to take an activated value, a_i is an activation probability for the value V_i of the i-th element in said each learning pattern to take an activated value, I_j is a predetermined threshold for the j-th node, b is a constant coefficient, and j and i are integers.

4. The apparatus of claim 3, wherein the constant coefficient b is equal to 2.

5. The apparatus of claim 1, wherein the link weight value updating means updates the link weight values according to values of the elements of said each learning pattern, the learning amount δ

determined by the learning amount determination means, and activation probabilities for the values of the elements in said each learning pattern to take activated values.

6. The apparatus of claim 5, wherein the link weight value updating means updates a link weight value W_ji for a link between a j-th node and an i-th node according to the following formula:

##EQU12## where W_ji (new) is a link weight value after updating, W_ji (old) is a link weight value before updating, V_j is a value of a j-th element in said each learning pattern, V_i is a value of an i-th element in said each learning pattern, a_j is an activation probability for the value V_j of the j-th element in said each learning pattern to take an activated value, a_i is an activation probability for the value V_i of the i-th element in said each learning pattern to take an activated value, and j and i are integers.

7. The apparatus of claim 1, further comprising activation probability updating means for updating activation probabilities for values of the elements in said each learning pattern to take activated values according to the values of the elements in said each learning pattern entered by the pattern entering means and the learning amount δ

determined by the learning amount determination means.

8. The apparatus of claim 7, wherein the activation probability updating means updates an activation probability a_j for a value of a j-th element in said each learning pattern to take an activated value according to the following formula:

space="preserve" listing-type="equation">a.sub.j (new)←

(1-δ

)a.sub.j (old)+δ

V.sub.j

where a_j (new) is an activation probability after updating, a_j (old) is an activation probability before updating, V_j is the value of the j-th element in said each learning pattern, and j is an integer.

9. A method of operating an associative memory in a form of a neural network in which a plurality of nodes having activation values are connected by a plurality of links having link weight values, comprising the steps of:

(a) entering a plurality of learning patterns into the neural network sequentially, each learning pattern having a plurality of elements in correspondence with the nodes;

(b) calculating an energy E of said each learning pattern entered at the step (a);

(c) determining a learning amount δ

for said each learning pattern according to a difference between the energy E calculated at the step (b) and a predetermined reference energy level Eth, using the following equation;

space="preserve" listing-type="equation">δ

=g(E-Eth)

where g is an upper and lower bounded monotonically increasing function with g(0)=0;

(d) updating the link weight values of the links in the neural network according to said each learning pattern entered at the step (a) and the learning amount δ

determined at the step (c), so as to achieve a learning of the associative memory;

(e) presenting an input pattern having a plurality of elements in correspondence with the nodes to the neural network;

(f) updating the activation values of the nodes in the neural network according to the input pattern presented at the step (e) and the link weight values of the links in the neural network updated at the step (d); and

(g) operating the associative memory for recalling an appropriate pattern according to the activation values of the nodes in the neural network updated at the step (f), so as to achieve an association by the associative memory.

10. The method of claim 9, wherein the energy E of said each learning pattern is calculated at the step (b) according to values of the elements of said each learning pattern, the link weight values previously updated at the step (d), and activation probabilities for the values of the elements in said each learning pattern to take activated values.

11. The method of claim 10, wherein the energy E of said each learning pattern is calculated at the step (b) according to the following equation:

##EQU13## where k is a non-zero constant coefficient having a sign identical to that of the predetermined reference energy level Eth W_ji is a link weight value for a link between a j-th node and an i-th node, V_j is a value of a j-th element in said each learning pattern, V_i is a value of an i-th element in said each learning pattern, a_j is an activation probability for the value V_j of the j-th element in said each learning pattern to take an activated value, a_i is an activation probability for the value V_i of the i-th element in said each learning pattern to take an activated value, I_j is a predetermined threshold for the j-th node, b is a constant coefficient, and j and i are integers.

12. The method of claim 11, wherein the constant coefficient b is equal to 2.

13. The method of claim 9, wherein the link weight values are updated at the step (d) according to values of the elements of said each learning pattern, the learning amount δ

determined at the step (c), and activation probabilities for the values of the elements in said each learning pattern to take activated values.

14. The method of claim 13, wherein the link weight value W_ji for a link between a j-th node and an i-th node is updated at the step (d) according to the following formula:

##EQU14## where W_ji (new) is a link weight value after updating, W_ji (old) is a link weight value before updating, V_j is a value of a i-th element in said each learning pattern, a_j is an activation probability for the value V_j of the j-th element in said each learning pattern to take an activated value, a_i is an activation probability for the value V_i of the i-th element in said each learning pattern to take an activated value, and j and i are integers.

15. The method of claim 9, further comprising the step of:

(h) updating activation probabilities for values of the elements in said each learning pattern to take activated values according to the values of the elements in said each learning pattern entered at the step (a) and the learning amount δ

determined at the step (c).

16. The method of claim 15, wherein an activation probability a_j for a value of a j-th element in said each learning pattern to take an activated value is updated at the step (h) according to the following formula:

space="preserve" listing-type="equation">a.sub.j (new)←

(1-δ

)a.sub.j (old)+δ

V.sub.j

where a_j (new) is an activation probability after updating, a_j (old) is an activation probability before updating, V_j is the value of the j-th element in said each learning pattern, and j is an integer.

17. The method of claim 9, wherein the activation values of the nodes are updated at the step (f) according to values of the elements of the input pattern presented at the step (e), the link weight values updated at the step (d), and activation probabilities for the values of the elements in the input pattern to take activated values.

18. The method of claim 17, wherein the activation values of the nodes are updated at the step (f) according to the following formula:

##EQU15## where V_j '"'"' is an activation value of a j-th node, W_ji is a link weight value for a link between the j-th node and an i-th node, V_i '"'"' is a value of an i-th element in the input pattern, a_i is an activation probability for the value V_i '"'"' of the i-th element in the input pattern to take an activated value, I_j is a predetermined threshold for the j-th node, f is a threshold function defined by the following equation;

##EQU16## where ##EQU17## and j and i are integers.