Numeric encoding method and apparatus for neural networks

Claim 1

A method for encoding a binary input into a form applicable to input neurons of a neural network comprising the steps of: partitioning said input into groups of three bits, the bits of each three-bit group representing an octal value; and replacing each three-bit group with a group of ten bits corresponding to said octal value, said ten-bit group consisting of three logic ones and seven logic zeros, said three logic ones occupying adjacent bit positions with the least significant of said three logic ones occupying a bit position of significance corresponding to said octal value.

Claim 2

A method for encoding a binary input into a form applicable to input neurons of a neural network comprising the steps of: partitioning said input into groups of three bits, the bits of each three-bit group representing an octal value; and replacing each three-bit group with a group of ten bits corresponding to said octal value, said ten-bit group consisting of three logic ones and seven logic zeros, said three logic ones occupying adjacent bit positions with the least significant of said three logic ones occupying a bit position of significance corresponding to said octal value.

Claim 3

A method for encoding a binary input into a form applicable to input neurons of a neural network comprising the steps of: partitioning said input into groups of three bits, the bits of each three-bit group representing an octal value; replacing each three-bit group with a group of ten bits corresponding to said octal value, said ten-bit group consisting of three logic ones and seven logic zeros, said three logic ones occupying adjacent bit positions with the least significant of said three logic ones occupying a bit position of significance corresponding to said octal value; and concatenating said ten-bit groups, while maintaining a relationship therebetween corresponding to that between said three-bit groups respectively.

Claim 4

A method for encoding a binary input into a form applicable to input neurons of a neural network comprising the steps of: partitioning said input into groups of three bits, the bits of each three-bit group representing an octal value; replacing each three-bit group with a group of ten bits corresponding to said octal value, said ten-bit group consisting of three logic ones and seven logic zeros, said three logic ones occupying adjacent bit positions with the least significant of said three logic ones occupying a bit position of significance corresponding to said octal value; and concatenating said ten-bit groups, while maintaining a relationship therebetween corresponding to that between said three-bit groups respectively.

Claim 5

A method for decoding a digital output from a neural network into a binary output comprising the steps of: partitioning the digital output into ten-bit output segments; replacing each of said ten-bit output segments with a three-bit binary segment corresponding to a bracketed weighted average of the significances of logic ones present in said output segment; and concatenating said three-bit binary segments to form the binary output.

Claim 6

A method for decoding a digital output from a neural network into a binary output comprising the steps of: partitioning the digital output into ten-bit output segments; replacing each of said ten-bit output segments with a three-bit binary segment corresponding to a bracketed weighted average of the significances of logic ones present in said output segment; and concatenating said three-bit binary segments to form the binary output.

Claim 7

A method for decoding a digital output from a neural network into a binary output comprising the steps of: partitioning the digital output into groups of ten-bit output segments, each bit of each group having a bit position value of from zero to nine corresponding to its significance within its group and a bit weight value equal to its bit position value minus one; converting each of said ten-bit output segments into a three-bit binary segment, said binary segment being a bracketed average of the bit weight values of all bits of said output segment that consist of logic ones; and concatenating said three-bit binary segments to form the binary output.

Claim 8

A method for decoding a digital output from a neural network into a binary output comprising the steps of: partitioning the digital output into groups of ten-bit output segments, each bit of each group having a bit position value of from zero to nine corresponding to its significance within its group and a bit weight value equal to its bit position value minus one; converting each of said ten-bit output segments into a three-bit binary segment, said binary segment being a bracketed average of the bit weight values of all bits of said output segment that consist of logic ones; and concatenating said three-bit binary segments to form the binary output.

Claim 9

A method for decoding a digital output from a neural network into a binary output comprising the steps of: A. partitioning the digital output into groups of ten-bit output segments, each bit of each group having a bit position value of from zero to nine corresponding to its significance within its group and a bit weight value equal to its bit position value minus one; B. converting each of said ten-bit output segments into a three-bit binary segment, said converting step comprising the steps of:

Claim 10

A method for decoding a digital output from a neural network into a binary output comprising the steps of: A. partitioning the digital output into groups of ten-bit output segments, each bit of each group having a bit position value of from zero to nine corresponding to its significance within its group and a bit weight value equal to its bit position value minus one; B. converting each of said ten-bit output segments into a three-bit binary segment, said converting step comprising the steps of:

Claim 13

A method for processing numeric data with a neural network comprising the steps of: A. encoding a binary input into a form directly applicable to input neurons of the neural network, said encoding step comprising the steps of:

Claim 14

A method for processing numeric data with a neural network comprising the steps of: A. encoding a binary input into a form directly applicable to input neurons of the neural network, said encoding step comprising the steps of:

Claim 17

A method for processing numeric data with a neural network comprising the steps of: A. encoding a binary input into a form directly applicable to input neurons of the neural network, said encoding step comprising the steps of: 1. partitioning said input into input segments of N bits each, each of said input segments representing a binary value; 2. replacing each input segment with a code corresponding to said binary value and consisting of M logic ones and 2 N -1 logic zeros, said M logic ones occupying adjacent bit positions in said code with the least significant of said M logic ones occupying a bit position in said code of significance corresponding to said binary value; and 3. concatenating said codes to form an encoded input; B. applying said encoded input to the input neurons of the neural network and generating an encoded output from the neural network; and C. decoding said encoded output to form a binary output, said decoding step comprising the steps of: 1. partitioning said encoded output into output segments of 2 N +M-1 bits each, each bit of each segment having an integer bit position value of from zero to 2 N +M-2 corresponding to its significance within its output segment and a bit weight value equal to its bit position value minus (M-1)/2; 2. converting each output segment into an N-bit binary segment, said converting step comprising the steps of: a. summing said bit weight values of all bits of said output segment that consist of logic ones; b. dividing said summation by the number of bits of said output segment that consist of logic ones, if any, to form a quotient; c. rounding said quotient to a nearest integer; and d. bracketing said rounded quotient between values of zero and 2 N -1, forming an N-bit binary segment; and

Claim 18

A method for processing numeric data with a neural network comprising the steps of: A. encoding a binary input into a form directly applicable to input neurons of the neural network, said encoding step comprising the steps of: 1. partitioning said input into input segments of N bits each, each of said input segments representing a binary value; 2. replacing each input segment with a code corresponding to said binary value and consisting of M logic ones and 2 N -1 logic zeros, said M logic ones occupying adjacent bit positions in said code with the least significant of said M logic ones occupying a bit position in said code of significance corresponding to said binary value; and 3. concatenating said codes to form an encoded input; B. applying said encoded input to the input neurons of the neural network and generating an encoded output from the neural network; and C. decoding said encoded output to form a binary output, said decoding step comprising the steps of: 1. partitioning said encoded output into output segments of 2 N +M-1 bits each, each bit of each segment having an integer bit position value of from zero to 2 N +M-2 corresponding to its significance within its output segment and a bit weight value equal to its bit position value minus (M-1)/2; 2. converting each output segment into an N-bit binary segment, said converting step comprising the steps of: a. summing said bit weight values of all bits of said output segment that consist of logic ones; b. dividing said summation by the number of bits of said output segment that consist of logic ones, if any, to form a quotient; c. rounding said quotient to a nearest integer; and d. bracketing said rounded quotient between values of zero and 2 N -1, forming an N-bit binary segment; and

Claim 22

A method for encoding a binary input into a form applicable to input neurons of a neural network comprising the steps of: partitioning the input into input segments of N bits each, each of said input segments representing a binary value; replacing each N-bit segment with a code corresponding to said binary value and consisting of M logic ones and 2 N -1 logic zeros, said M logic ones occupying adjacent bit positions in said code with the least significant of said M logic ones occupying a bit position in said code of significance corresponding to said binary value; and concatenating said codes to form an encoded input.

Claim 23

A method for encoding a binary input into a form applicable to input neurons of a neural network comprising the steps of: partitioning the input into input segments of N bits each, each of said input segments representing a binary value; replacing each N-bit segment with a code corresponding to said binary value and consisting of M logic ones and 2 N -1 logic zeros, said M logic ones occupying adjacent bit positions in said code with the least significant of said M logic ones occupying a bit position in said code of significance corresponding to said binary value; and concatenating said codes to form an encoded input.

Claim 24

A method for decoding a digital output from a neural network into a binary output comprising the steps of: partitioning the digital output into output segments of 2 N +M-1 bits each, each bit cf each segment having an integer bit position value of from zero to 2 N +M-2 corresponding to its significance within its segment and a bit weight value equal to its bit position value minus (M-1)/2; converting each output segment into a N-bit binary segment, said binary segment being a bracketed average of the bit weight values of all bits of said output segment that consist of logic ones; and concatenating said N-bit binary segments to form the binary output.

Claim 25

A method for decoding a digital output from a neural network into a binary output comprising the steps of: partitioning the digital output into output segments of 2 N +M-1 bits each, each bit cf each segment having an integer bit position value of from zero to 2 N +M-2 corresponding to its significance within its segment and a bit weight value equal to its bit position value minus (M-1)/2; converting each output segment into a N-bit binary segment, said binary segment being a bracketed average of the bit weight values of all bits of said output segment that consist of logic ones; and concatenating said N-bit binary segments to form the binary output.

Claim 26

A method for decoding a digital output from a neural network into a binary output comprising the steps of: A. partitioning the digital output into output segments of 2 N +M-1 bits each, each bit of each segment having an integer bit position value of from zero to 2 N +M-2 corresponding to its significance within its segment and a bit weight value equal to its bit position value minus (M-1)/2; B. converting each output segment into an N-bit binary segment, said converting step comprising the steps of: 1. summing said bit weight values of all bits of said output segment that consist of logic ones;

Claim 27

A method for decoding a digital output from a neural network into a binary output comprising the steps of: A. partitioning the digital output into output segments of 2 N +M-1 bits each, each bit of each segment having an integer bit position value of from zero to 2 N +M-2 corresponding to its significance within its segment and a bit weight value equal to its bit position value minus (M-1)/2; B. converting each output segment into an N-bit binary segment, said converting step comprising the steps of: 1. summing said bit weight values of all bits of said output segment that consist of logic ones;