Neural cortex

• US 20040034609A1
• Filed: 04/01/2003
• Published: 02/19/2004
• Est. Priority Date: 11/30/2000
• Status: Active Grant

• Current Assignee: Alexei Mikhailov, Yang Ming Pok
• Original Assignee: Alexei Mikhailov, Yang Ming Pok
• Inventors: Mikhailov, Alexei, Pok, Yang Ming
• Granted Patent: US 7,305,370 B2
• Time in Patent Office: Days
• US Class Current: 706/20
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A neural network system includes a random access memory (RAM); and an index-based weightless neural network with a columnar topography; wherein patterns of binary connections and values of output nodes'"'"' activities are stored in the RAM. Information is processed by pattern recognition using the neural network by storing a plurality of output patterns to be recognised in a pattern index; accepting an input pattern and dividing the input pattern into a plurality of components; and processing each component according to the pattern index to identify a recognised output pattern corresponding to the input pattern.

Claims

1. A neural network system including:

(a) a random access memory (RAM); and

(b) an index-based weightless neural network with a columnar topography;

wherein patterns of binary connections and values of output nodes'"'"' activities are stored in the RAM.

2. A neural network system according to claim 1 wherein the system comprises a computer hardware component.

3. A neural network system according to claim 1 or claim 2 wherein systematic expansion is achieved by increasing the size of the RAM.

4. A neural network system according to any one of claims 1 to 3 wherein the neural network is trained by writing of data into the RAM, and network topology emerges during the training.

5. A neural network system according to any one of claims 1 to 4 wherein performance is adjustable by changing decomposition style of input data, and thereby changing dynamic range of input components.

6. A neural network system according to any one of claims 1 to 5 wherein input components address a single common index.

7. A method of processing information by pattern recognition using a neural network including the steps of—

(a) storing a plurality of output patterns to be recognised in a pattern index;

(b) accepting an input pattern and dividing the input pattern into a plurality of components;

(c) processing each component according to the pattern index to identify a recognised output pattern corresponding to the input pattern.

8. A method according to claim 7 wherein each output pattern is divided into a plurality of recognised components with each recognised component being stored in the pattern index for recognition.

9. A method according to claim 8 wherein the index consists of columns with each column corresponding to one or more recognised components.

10. A method according to claim 9 wherein the index is divided into a number of columns which is equal to or less than the number of recognised components.

11. A method according to claim 9 wherein the index is divided into a number of columns which is equal to the number of recognised components.

12. A method according to any one of claims 8 to 10 wherein each input component is compared to the corresponding recognised component column and a score is allocated to one or more recognised components.

13. A method according to claim 12 wherein the score for each recognised component of a pattern is added and the recognised pattern with the highest score is identified as the output pattern.