Large memory storage and retrieval (LAMSTAR) network
First Claim
1. A network system for processing an input word comprising an associated set of subword category subwords, the input word being an input observation of a monitored system as of a specific time period, said system comprising:
- means for storing the input word;
a SOM subsystem comprising a plurality of self-organizing map modules (SOM), each said SOM module comprising a plurality of memory cells for selectively storing, as stored subwords, the category subwords of the input word, said system further comprising;
means for comparing and correlating said category subwords of the input word to determine a winning match between one of the subwords of the input word and a finite predetermined number of the memory cells for subwords stored within said SOM modules;
means for choosing a selected one of the memory cells in a selected one of said SOM modules responsive to said winning match;
a plurality of sets of subword category correlating links and of SOM memory cell correlating links, each said correlating link comprising a memory cell link weight value for linking respective associated ones of selected ones of the memory cells between respective associated selected ones of the SOM modules;
a task evaluator subsystem comprised of means for establishing threshold values for said correlating links;
said system further comprising;
means for adjusting said memory cell link weight values responsive to the input subwords and to the SOM memory cells, said memory cell link weight values being increased upon said winning match, said memory cell link weight values being decreased when no said winning match is established; and
means for storing a subword in one of the SOM memory cells when no winning match is determined after a predetermined number of memory cells in the SOM modules are compared.
1 Assignment
0 Petitions
Accused Products
Abstract
A network system analyzes input words for the search and retrieval of pertinent information. The novel system then selects a module of a self organizing map (SOM) which contains the same dimension of classification as a selected input word and where neurons are interconnected horizontally (between modules) and vertically (at input and inside a module) by arrays of link weights. The system then determines what nodes or processing units within the SOM will be activated and subsequently compared to the selected input word. Feedback is utilized via a punishment/reward scheme to adjust the link weights so that the system learns the best paths and/or methods to create acceptable decisions or outputs.
78 Citations
55 Claims
-
1. A network system for processing an input word comprising an associated set of subword category subwords, the input word being an input observation of a monitored system as of a specific time period, said system comprising:
-
means for storing the input word; a SOM subsystem comprising a plurality of self-organizing map modules (SOM), each said SOM module comprising a plurality of memory cells for selectively storing, as stored subwords, the category subwords of the input word, said system further comprising; means for comparing and correlating said category subwords of the input word to determine a winning match between one of the subwords of the input word and a finite predetermined number of the memory cells for subwords stored within said SOM modules; means for choosing a selected one of the memory cells in a selected one of said SOM modules responsive to said winning match; a plurality of sets of subword category correlating links and of SOM memory cell correlating links, each said correlating link comprising a memory cell link weight value for linking respective associated ones of selected ones of the memory cells between respective associated selected ones of the SOM modules; a task evaluator subsystem comprised of means for establishing threshold values for said correlating links; said system further comprising; means for adjusting said memory cell link weight values responsive to the input subwords and to the SOM memory cells, said memory cell link weight values being increased upon said winning match, said memory cell link weight values being decreased when no said winning match is established; and means for storing a subword in one of the SOM memory cells when no winning match is determined after a predetermined number of memory cells in the SOM modules are compared. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22)
-
-
23. A learning network system for providing a diagnostic result output responsive to an input word comprised of a plurality of input subwords, said system comprising:
-
control apparatus for determining link weight values for each of the input subwords; correlative link memory for selectively storing and outputting the link weight value for a plurality of subword categories, one said value corresponding to each said subword category, wherein each of the plurality of input subwords is assigned to one of the plurality of subword categories, said correlative link memory further comprising initialization link correlative memory, vertical link correlative memory, and horizontal link correlative memory; SOM memory comprising a plurality of self-organizing map modules (SOM), each said module comprising a plurality of memory cells, each said module providing for storing a plurality of the input subwords associated with a single one of the assigned subword categories; means for processing the input word responsive to the initialization link correlative memory to select a respective one of the input subwords as a selected subword to be processed, wherein the selected input subword is coupled for correlation to a corresponding selected one of the SOM modules; and means for correlating the selected subword to at least one of a predefined plurality of the memory cells of the selected one of the SOM modules responsive to the correlative link memory to provide a result output; wherein the control apparatus is responsive to the result output to modify the respective link weight values as stored in the correlative link memory to provide learning and forgetting. - View Dependent Claims (24, 25, 26, 27)
-
-
28. A neural network system for providing an output decision to a user, said system responsive to an input word, said input word comprising a plurality of input subwords, each said input subword having an associated subword category, said neural network system comprising:
-
apparatus for initially determining a link weight value for each subword category; a correlative link memory for selectively storing and outputting the link weight values for each of the subword categories of the input word, said correlative link memory comprising initialization link memory, vertical link memory, and horizontal link memory; a plurality of self-organizing map modules (SOM), wherein each of the SOM modules is comprised of a plurality of memory cells, each of said SOM modules providing for storage of input subwords relating to a common one of the subword categories, wherein there is a SOM module for each of the plurality of input subword categories, wherein each of the plurality of SOM modules communicates with the correlative link memory; means for selecting a particular one of the input subwords for processing responsive to correlating the input subwords against the link weight values stored in the initialization link memory; means for selecting a particular one of the SOM modules responsive to correlating the particular one of the input subwords against the link weight values in the vertical link memory; means for selecting a particular one of the memory cells within the particular one of the SOM modules responsive to correlating the particular one of the input subwords against the link weight values stored in the horizontal link memory; and adjustment means for modifying the link weight values stored in the correlative link memory responsive to learning and forgetting based on the correlating of the input subwords against the correlative link memory. - View Dependent Claims (29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48)
-
-
49. A neural network system for providing an output decision to a user, said system responsive to an input word, said input word comprising a plurality of input subwords, each said input subword having an associated subword category, said neural network system comprising:
- a correlative link memory for storing and selectively outputting correlative link weights for each of the subword categories of the input word, said correlative link memory comprising vertical link memory, and horizontal link memory;
a task evaluator subsystem comprising means for establishing associated vertical and horizontal link thresholds for each of said respective vertical and horizontal link memories; SOM memory comprising a plurality of self-organizing map modules (SOM), wherein each of the SOM modules is comprised of a plurality of memory cells, each of said SOM modules providing for storage of input subwords relating to a common one of the subword categories, wherein there is a SOM module for each of the plurality of input subword categories, wherein each of the plurality of SOM modules is associated with a respective portion of the correlative link memory; means for selecting a particular one of the SOM modules responsive to correlating a selected one of the input subwords with said respective portion of the vertical link memory in accordance with said vertical link thresholds; and means for selecting a particular one of the memory cells within the particular one of the SOM modules responsive to correlating the selected one of the input subwords with said respective portion of the horizontal link memory in accordance with said horizontal link thresholds. - View Dependent Claims (50, 51, 52, 53)
- a correlative link memory for storing and selectively outputting correlative link weights for each of the subword categories of the input word, said correlative link memory comprising vertical link memory, and horizontal link memory;
-
54. A system to determine a result output responsive to processing an input word comprising a plurality of input subwords each having a respective associate subword category, the system comprising:
-
SOM memory comprising a plurality of SOM modules each comprised of a plurality of memory cells, for selectively storing and outputting indexed knowledge data; correlative link memory providing for selective storage and output of link weight values for each of the subword categories, comprising; vertical link memory providing for storage and output of the link weight values for each of the subword categories for selection of a respective one of the SOM modules; horizontal link weight memory for storage and output of the link weight values for each of the subword categories for selection of selected ones of the memory cells within the respective one of the SOM modules; first selection apparatus for selecting one of the input subwords as a selected subword; a second selection apparatus, responsive to the selected subword and the correlative link memory, for selecting a particular one of the SOM modules responsive to the correlation of the vertical link weight values to the selected subword; third selection apparatus for selecting at least a particular one of the memory cells for correlation within the particular one of the SOM modules against the selected subword, responsive to the selected subword and the horizontal link memory; result output apparatus for providing success and failure feedback responsive to the correlations of the second and third selection apparatus; and adjustment apparatus for selectively increasing and decreasing respective ones of the link weight values as stored in the correlative memory, responsive to the success and failure feedback. - View Dependent Claims (55)
-
Specification