SYSTEM AND METHOD FOR MODELING THE NEOCORTEX AND USES THEREFOR

• US 20080152217A1
• Filed: 05/15/2007
• Published: 06/26/2008
• Est. Priority Date: 05/16/2006
• Status: Active Grant

• Current Assignee: General Manifolds LLC
• Original Assignee: Douglas S. Greer
• Inventors: GREER, DOUGLAS S.
• Granted Patent: US 7,805,386 B2
• Time in Patent Office: Days
• US Class Current: 382/155
• CPC Class Codes: G06F 18/21355 nonlinear criteria, e.g. embedding a manifold in a Euclidean space
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06V 10/7715 Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods

Abstract

A processor architecture for a learning machine is presented which uses a massive array of processing elements having local, recurrent connections to form global associations between functions defined on manifolds. Associations between these functions provide the basis for learning cause-and-effect relationships involving vision, audition, tactile sensation and kinetic motion. Two arbitrary images hold each other in place in a manifold association processor and form the basis of short-term memory.

Claims

1. A method for generating an encoded output signal from an input signal having a structure comprising a plurality of discrete signal units, the method comprising:

inputting the input signal into a computer processor having software stored therein for executing a plurality of processing elements, wherein each processing element comprises a logic manifold association processor having an input and an output, and wherein each processing element generates a single output signal;

forming a recursive loop between a pair of processing elements, wherein the output of each pair of processing elements is fed back into the input of the other processing element, and wherein at least one of the pair of processing elements further receives an element input comprising a small subset of the plurality of discrete signal units of the input signal;

training a plurality of pairs of processing elements to store a plurality of reciprocal signal pairs, wherein when a first signal of a reciprocal signal pair of the plurality is input into a first processing element, the first processing element outputs a second signal of the reciprocal signal pair, and when the second signal of the reciprocal signal pair is input into a second processing element, the second processing element outputs the first signal, so that the reciprocal signal pair is locked within the pair of processing elements;

wherein when the at least one of the pair of processing elements that receives the small subset of discrete signal units receives the input signal units, it forces the two processing elements to generate an encoded output associated with the subset of input signal units.

2. The method of claim 1, wherein the input signal comprises a plurality of the small subsets of discrete signal units, each corresponding to one or more locations within the structure, and wherein each small subset is input into a different pair of processing elements within an array of processing elements in which the different pairs of processing elements are at least partially interconnected, and further comprising combining the outputs of the array of processing elements to generate the encoded output signal.

3. The method of claim 2, wherein the input signal comprises a first image and the structure comprises a pixel array, and wherein each subset of discrete signal units comprises one or more pixels.

4. The method of claim 3, wherein the encoded output signal comprises a second image different from the first image.

5. The method of claim 2, wherein each small subset of discrete signal units corresponds to a local region within the structure.

6. The method of claim 1, wherein multiple input signals are input into the processing elements with each of the multiple input signals topographically aligned with other input signals, so that a same region is used within each of the input signals.

7. The method of claim 1, wherein the input signal comprises a linearly transformed space and the encoded output signal comprises an image.

8. The method of claim 1, further comprising:

applying an input control mask to the input signal to control where an associate occurs.

9. The method of claim 8, wherein the input control mask varies with time.

10. The method of claim 1, further comprising:

applying an output control mask for blanking the outputs of one or more processing elements to prevent its input into other processing elements.

11. The method of claim 10, wherein the output control mask varies with time.

12. The method of claim 1, further comprising applying at least one convolution kernel for computing linear transforms over small regions of the input signal structure.

13. The method of claim 12, wherein the at least one convolution kernel comprises a plurality of kernels having different shapes and sizes for computing a frame operator.

14. The method of claim 1, wherein the input signal is transformed into an linearly transformed image space and the processing elements operate in linearly transformed image space.

15. The method of claim 1, wherein the plurality of processing elements comprises a neural network.

16. The method of claim 1, wherein the output signals of the pair of processing elements comprise coefficients of a frame operator, the frame coefficients including redundant information, wherein the method further comprises using a reproducing kernel and the redundant information to reduce noise in the output signals.

17. A method for extracting knowledge from an input signal having a structure comprising a plurality of discrete signal units, the method comprising:

transforming the input signal to a higher-dimensional space using a spectral transform to generate an input spectral representation;

using an array of processing elements executed by a computer processor, multiplying values of the input spectral representation to produce an associated value, wherein each processing element of the array comprises a different function;

mapping the associated value of each processing element by a nonlinear transfer function to generate a processing element output;

constructing an estimate of an output spectral representation from the array of mapped processing element outputs;

applying a linear transform defined by a reproducing kernel to produce an output signal comprising knowledge extracted from the input signal; and

displaying the output signal at an output device.

18. The method of claim 17, further comprising, prior to the step of multiplying, transforming the input spectral representation in each processing element by computing the pointwise difference between the input signal and a predetermined pattern;

and

mapping the pointwise difference by a basis function that has its maximal value at zero and progressively smaller values as its input goes to plus or minus infinity.

19. The method of claim 17, wherein the processing element output is recursively fed back and used as the input signal.

20. The method of claim 17, wherein the array of processing elements is extended to accept multiple input signals by:

aligning the multiple input signals topologically; and

extending the inputs in each processing element by using the same or similar processing element regions in each of the multiple input signals.

21. The method of claim 17, further comprising:

applying an input control mask to the input signal to control where an associate occurs.

22. The method of claim 21, wherein the input control mask varies with time.

23. The method of claim 17, further comprising:

applying an output control mask for blanking the outputs of one or more processing elements to prevent its input into other processing elements.

24. The method of claim 23, wherein the output control mask varies with time.

25. The method of claim 17, further comprising applying a convolution kernel for computing linear transforms over small regions of the input signal structure.

26. The method of claim 25, wherein the at least one convolution kernel comprises a plurality of kernels having different shapes and sizes for computing a frame operator.

27. The method of claim 17, wherein the output signals of the pair of processing elements comprise coefficients of a frame operator, the frame coefficients including redundant information, wherein the method further comprises using a reproducing kernel and the redundant information to reduce noise in the output signals.

28. A system for encoding or decoding an input signal, the input signal having a signal structure, the system comprising:

a computer having stored therein software for executing a plurality of manifold association processors and for receiving the input signal, wherein each manifold association processor comprises;

a processor input and a processor output and wherein each processor generates a single output signal;

a recursive loop between a pair of processors, wherein the output of each pair of processors is fed back into the input of the other processor, and wherein at least one of the pair of processors further receives an element input comprising a small subset of discrete signal units of the input signal, wherein the processors are trained to store a plurality of reciprocal signal pairs, wherein when a first signal of a reciprocal signal pair of the plurality is input into a first processing element, the first processing element outputs a second signal of the reciprocal signal pair, and when the second signal of the reciprocal signal pair is input into a second processing element, the second processing element outputs the first signal, so that the reciprocal signal pair is locked within the pair of processing elements, wherein when the at least one of the pair of processing elements that receives the small subset of discrete signal units receives the input signal units, it forces the two processing elements to generate an encoded output associated with the subset of input signal units; and

means for combining the processor outputs of all the individual processors to form the encoded output signal corresponding to the input signal.

29. The method of claim 28, wherein the input signal comprises a first image and the structure comprises a pixel array, and wherein each subset of discrete signal units comprises one or more pixels.

30. The method of claim 29, wherein the encoded output signal comprises a second image different from the first image.

31. The method of claim 29, wherein each small subset of discrete signal units corresponds to a local region within the structure.

32. The method of claim 28, wherein multiple input signals are input into the processing elements with each of the multiple input signals topographically aligned with other input signals, so that a same region is used within each of the input signals

33. The method of claim 28, wherein the input signal comprises a linearly transformed space and the encoded output signal comprises an image.