US 10,242,313 B2Granted

Joint proximity association template for neural networks

OwnerJames LaRue, Denise LaRueinventors

Life of this patent

Estimating the term…

Loading patent…

Abstract

A technical solution is described for implementing a computer-executed system of association memory matrices to replace the proximal layers of a convolutional neural network (CNN). An example method includes configuring one Associative Memory Matrix (AMM) for each configured layer in the CNN. This one-to-one conversion method motivates the name to the product: the Joint Proximity Association Template (JPAT) for Neural Networks. The invention is a numerically stable soft-ware based implementation that (1) reduces the long training times, (2) reduces the execution time, and (3) produces bidirectional intra-layer connections and potentially, inter-layer connections as well. The method further includes, potentially, forming a single AMM, from the multiple AMMs corresponding to the multiple and proximal layers of the CNN, in anticipation of the well-known Universal Approximation Theorem.

Claims3

Dependent claims

A computer implemented Joint Proximity Association Template (JPAT) neural network method for implementing a plurality of intra-connected associative memory matrices, based on a set of hierarchical feature detection layers internal to a convolutional neural network based image classifier, the method comprising the steps of:

(a) Configure a computer for inputting a plurality of patterns of images into said convolutional neural network pre-configured to classify said images into to one of N classes of patterns of images;
(b) Configure processor of said computer to record into memory an output vector from each said feature detection layer of said network, corresponding to each said input pattern processed through said network;
(c) Retrieve from memory one input pattern stored in vector form and retrieve said input pattern's corresponding output vector from the first said detection layer;
process the multiplication of said input vector by the transpose of said output vector to form a matrix of numbers and store into memory;
(d) Repeat processing step (c) for remaining said input vectors and corresponding first said detection layer output vectors to form and store into memory additional matrices;
(e) Retrieve from memory the matrices formed in steps (c) and (d) and implement said configured processor to add together said matrices to produce a single matrix of numbers;
store this single matrix in memory;
(f) Invoke processor to extract from memory said single matrix and calculate a singular value decomposition (SVD) of said single matrix and store in computer memory a matrix of left-singular vectors, a diagonal matrix containing singular values on the diagonal, and a matrix of right-singular vectors;
(g) Retrieve from memory the pre-configured number N;
extract from said left-singular vectors a submatrix of N vectors which correspond to the N highest singular values along said diagonal matrix;
extract from said right-singular vectors a submatrix of N vectors which correspond to same N highest singular values along said diagonal matrix;
invoke processor to multiply said submatrix of left singular vectors by the transpose of said submatrix of right-singular vectors to produce an associative memory matrix (AMM);
(h) Repeat steps (c)-(g) to calculate an associative memory matrix (AMM) from each of the remaining consecutive layers;
(i) Store in computer memory the plurality of said associative memory matrices to form an array of intra-connected associative memory matrices corresponding to said feature detection layers of said convolutional neural network;
(j) Configure a computer for inputting a plurality of patterns of images into said array of associative memory matrices to process and classify said images into to one of N classes of patterns of images.

12 more elements

A system facilitating the calculation of a Joint Proximity Association Template (JPAT) neural network that implements a plurality of intra-connected associative memory matrices, based on a set of hierarchical feature detection layers internal to a convolutional neural network based image classifier, wherein a processor is configured to:

(a) Configure a computer system for inputting a plurality of patterns of images into said convolutional neural network pre-configured to classify said images into to one of N classes of patterns of images;
(b) Configure system processor of said computer system to record into memory an output vector from each said feature detection layer of said network, corresponding to each said input pattern processed through said network;
(c) Retrieve from system memory one input pattern stored in vector form and retrieve said input pattern's corresponding output vector from the first said detection layer;
process the multiplication of said input vector by the transpose of said output vector to form a matrix of numbers and store into system memory;
(d) Repeat processing step (c) for remaining said input vectors and corresponding first said detection layer output vectors to form and store into memory additional matrices;
(e) Retrieve from system memory the matrices formed in steps (c) and (d) and implement said configured processor to add together said matrices to produce a single matrix of numbers;
store this single matrix into system memory;
(f) Invoke system processor to extract from memory said single matrix and calculate a singular value decomposition (SVD) of said single matrix and store in computer memory a matrix of left-singular vectors, a diagonal matrix containing singular values on the diagonal, and a matrix of right-singular vectors;
(g) Retrieve from system memory the pre-configured number N;
extract from said left-singular vectors a submatrix of N vectors which correspond to the N highest singular values along said diagonal matrix;
extract from said right-singular vectors a submatrix of N vectors which correspond to same N highest singular values along said diagonal matrix;
invoke system processor to multiply said submatrix of left singular vectors by the transpose of said submatrix of right-singular vectors to produce an associative memory matrix (AMM);
(h) Repeat steps (c)-(g) to calculate an associative memory matrix (AMM) from each of the remaining consecutive layers;
(i) Store in computer system memory the plurality of said associative memory matrices to form an array of intra-connected associative memory matrices corresponding to said feature detection layers of said convolutional neural network;
(j) Configure a computer system for inputting a plurality of patterns of images into said array of associative memory matrices to process and classify said images into to one of N classes of patterns of images.

12 more elements

A computer program product for calculating a Joint Proximity Association Template (JPAT) neural network product that implements a plurality of intra-connected associative memory matrices, based on a set of hierarchical feature detection layers internal to a convolutional neural network based image classifier, said computer program product comprising a non-transitory computer readable storage medium, the non-transitory computer readable storage medium comprising computer executable instructions, wherein the non-transitory computer readable storage medium comprises instructions to:

(a) Configure a computer product for inputting a plurality of patterns of images into said convolutional neural network pre-configured to classify said images into to one of N classes of patterns of images;
(b) Configure processor of said computer product wherein the non-transitory computer readable storage medium further comprises instructions to record into memory an output vector from each said feature detection layer of said network, corresponding to each said input pattern processed through said network;
(c) Configure the non-transitory computer program product to retrieve from system memory one input pattern stored in vector form and retrieve said input pattern's corresponding output vector from the first said detection layer;
process the multiplication of said input vector by the transpose of said output vector to form a matrix of numbers and store into memory;
(d) Repeat processing step (c) for remaining said input vectors and corresponding first said detection layer output vectors to form and store into system memory additional matrices;
(e) Retrieve from memory the matrices formed in steps (c) and (d) and implement said configured processor to add together said matrices to produce a single matrix of numbers;
store this single matrix in memory;
(f) Invoke the non-transitory computer readable storage medium processor to extract from system memory said single matrix and calculate a singular value decomposition (SVD) of said single matrix and store in computer memory a matrix of left-singular vectors, a diagonal matrix containing singular values on the diagonal, and a matrix of right-singular vectors;
(g) Retrieve from system memory the pre-configured number N;
extract from said left-singular vectors a submatrix of N vectors which correspond to the N highest singular values along said diagonal matrix;
extract from said right-singular vectors a submatrix of N vectors which correspond to same N highest singular values along said diagonal matrix;
invoke processor to multiply said submatrix of left singular vectors by the transpose of said submatrix of right-singular vectors to produce an associative memory matrix (AMM);
(h) Repeat steps (c)-(g) to calculate an associative memory matrix (AMM) from each of the remaining consecutive layers;
(i) Configure said non-transitory computer readable storage medium to store into memory the plurality of said associative memory matrices to form an array of intra-connected associative memory matrices corresponding to said feature detection layers of said convolutional neural network;
(j) Configure a non-transitory computer system product for inputting a plurality of patterns of images into said array of associative memory matrices to process and classify said images into to one of N classes of patterns of images.

12 more elements

Dependencies are inferred from the claim text.