Image analysis – Histogram processing – For setting a threshold
Patent
1993-10-20
1996-05-07
Moore, David K.
Image analysis
Histogram processing
For setting a threshold
395 21, 395 24, 382156, G06F 1580, G06F 1518
Patent
active
055154778
DESCRIPTION:
BRIEF SUMMARY
TECHNICAL FIELD
The present invention relates to adaptive information processing systems, and in particular to associative memories utilizing confidence-mediated associations, and especially neural network systems comprising an auto-organizational apparatus and processes for dynamically mapping an input onto a semantically congruous and contemporaneously-valid, learned response.
BACKGROUND OF ART
Broadly speaking, an associative memory system is one in which stimulus/response pairs of information are stored in such a way that the introduction of a stimulus pattern results in the recall of a memory associated response. Memory systems of this type have a very broad range of potential applications including, for example, logical operations management, pattern recognition, and image interpolation.
Traditional associative processes, such as those that are often used in artificial intelligence applications, are dependent on explicitly predefined rule sets that are externally impressed on an associative memory. Expert systems are examples of such traditional architectures. Such expert systems are rules-based paradigms that are managed by an inferential engine. These follow an orthodox von Neumann approach by providing a deterministic software/hardware relationship that follows a series of pre-declared relationships and sequential instructions formatted as predetermined sets of IF--THEN statements. They are inherently limited to those associations that are expressly pre-ordained or are expressly permitted to be logically deduced by preestablished inferential rules. There is no intrinsic adaptive capability in these processes. In consequence there is no dynamic responsiveness to changing environments or, more generally, any ability to develop a set of input-appropriate responses in the absence of an impressed set of applicable rules specifically intended to deal with a changed or changing or otherwise unknown environment. Moreover, as with any purely heuristic programming, the more complex the application, the greater the number of rules that are required, and the proportionately longer the processing time required to deal with those rules. There is a general acceptance that these short comings limit the practical usefulness of pre-defined-rules-based approaches to associative memory systems.
Neural networks, on the other hand, generate their own rules of association through a learning process that draws on the networks exposure to either supervised or unsupervised input data samples drawn from a statistical universe. These systems have, to various degrees, some ability to make generalizations about that universe as a whole, based on the input sampling.
Neural networks are associative memory systems comprising strategic organizations, (architectures), of processing elements. Individually, these elements are each analogous to an individual neuron in a biological system. Individual processing elements have a plurality of inputs, which are functionally analogous to the dendritic processes of a neuron cell. As such, these elements are conditioned in accordance with a paradigm over the course of an ongoing learning process, to dynamically assign and assert a ceratin "weight", based on the current state of the systems knowledge, to the respective inputs. The associative "weights" form the data that is stored in the associative memory of the system. Digital computer implementations of neural networks typically employ numerical methodologies to realize the desired associative recall of stimulus-appropriate responses through weighted summation of the inputs in a digital computing environment. These virtual networks take advantage of the current commercial availability of von Neumann machines, which while inherently deterministic, are nevertheless capable of being used to advantages attached to stochastic architectures in neural network hardware implementations.
An early forerunner to modern neural networks, howsoever they may now be implemented, was an actual hardware device that came to be known as the Perceptron. This
REFERENCES:
patent: 4979126 (1990-12-01), Pao et al.
patent: 5040230 (1991-08-01), Takatori et al.
patent: 5050095 (1991-09-01), Samad
patent: 5214745 (1993-05-01), Sutherland
Grant et al, "Synthesis of a class of artificial neural network using a CMOS current mode building block approach"; IEEE Colloquim on `Advances in Analogue VLSI`, p. 8/1-10, 14 May 1991.
Tsai et al, "An associative memory knowledge base for diagnostic and high level control functions"; Proceedings of the 2nd international IEEE conference on tools for artificial intelligence, p. 283-8, 6-9 Nov. 1990.
Specht, "Probabalistic nerual networks for classification, mapping, or associative memory"; IEEE International Conference on Neural Networks, p. 525-32 vol. 1, 24-27 Jul. 1988.
Hafiz Tariq Rafiq
Moore David K.
LandOfFree
Neural networks does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Neural networks, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Neural networks will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1233981