Neuron unit, neural network and signal processing method

Claim 2

A signal processing system, comprising: N neural networks which are coupled to carry out signal processing on a plurality of input signals respectively having N bits, where N is an integer, the N neural networks receiving corresponding bits of the input signals, each of the N neural networks including: A) a plurality of neuron units which are coupled to form a hierarchical structure which has a plurality of layers, each of the neuron units simultaneously processing a plurality of binary input signals and outputting an output signal which is indicative of a result of the processing, the neuron units including: a) input line means for receiving the input signals, the input signals being binary pulse-density-modulated signals; b) forward process means, coupled to the input line means, and including: 1) supplying means, made up of digital circuits, for supplying weight functions; and 2) operation means, made up of digital circuits, for carrying out an operation on each of the input signals using one of the weight functions and for outputting the output signal, the output signal being a binary pulse-density-modulated signal; and c) self-learning means, coupled to the forward process means, and including: 1) generating means for generating new weight functions based on errors between the output signal of the forward process means and teaching signals; and 2) varying means for varying the weight functions supplied by the supplying means of the forward process means to the new weight functions generated by the generating means; and B) a plurality of signal lines coupling outputs of arbitrary neuron units in one layer of the hierarchical structure to inputs of arbitrary neuron units in another layer of the hierarchical structure.

Claim 3

The system of claim 1, wherein: the errors are pulse-density-modulated.

Claim 4

A neuron unit for processing a plurality of input signals and for outputting an output signal which is indicative of a result of the processing, the neuron unit comprising: a) input line means for receiving the input signals, the input signals being binary pulse-density-modulated signals; b) forward process means, coupled to the input line means, and including: 1) supplying means, made up of digital circuits, for supplying weight functions; and 2) operation means, made up of digital circuits, for carrying out an operation on each of the input signals using one of the weight functions and for outputting the output signal, the output signal being a binary pulse-density-modulated signal; and c) self-learning means, coupled to the forward process means, and including: 1) generating means for generating new weight functions based on errors between the output signal of the forward process means and teaching signals; and 2) varying means for varying the weight functions supplied by the supplying means of the forward process means to the new weight functions generated by the generating means.

Claim 5

A neuron unit for processing a plurality of input signals and for outputting an output signal which is indicative of a result of the processing, the neuron unit comprising: a) input line means for receiving the input signals, the input signals being binary pulse-density-modulated signals; b) forward process means, coupled to the input line means, and including: 1) supplying means, made up of digital circuits, for supplying weight functions; and 2) operation means, made up of digital circuits, for carrying out an operation on each of the input signals using one of the weight functions and for outputting the output signal, the output signal being a binary pulse-density-modulated signal; and c) self-learning means, coupled to the forward process means, and including: 1) generating means for generating new weight functions based on errors between the output signal of the forward process means and teaching signals; and 2) varying means for varying the weight functions supplied by the supplying means of the forward process means to the new weight functions generated by the generating means.

Claim 6

The neuron unit as claimed in claim 3, wherein said input line means includes a plurality of first input lines for receiving first binary input signals which undergo transitions with time and a plurality of second input lines for receiving second binary input signals which undergo transitions with time, said supplying means includes first and second memory means for storing the weight functions, and said operation means includes: first gate means for successively obtaining a logical product of one of said first binary input signals received from said first input lines and a corresponding one of the weight functions read out from said first memory means for each of said first binary input signals; second gate means for successively obtaining a logical product of one of said second binary input signals received from said second input lines and a corresponding one of the weight functions read out from said second memory means for each of said second binary input signals; third gate means for obtaining a logical sum of logical products output from said first gate means; fourth gate means for obtaining a logical sum of logical products output from said second gate means; and output means including an inverter for inverting the logical sum output from said fourth gate means and a gate for obtaining one of a logical product and a logical sum of the logical sum output from said third gate means and an inverted logical sum output from said inverter, said gate outputting the output signal of the neuron unit.

Claim 7

The neuron unit as claimed in claim 4, wherein said first and second binary input signals respectively describe a signal quantity in a form of a pulse density, and said first and second memory means stores the weight functions which respectively describe a signal quantity in a form of a pulse density.

Claim 8

The neuron unit as claimed in claim 5, wherein each weight function has a pulse density defined by a number of first values and second values within a predetermined time, the first values and the second values are arranged at random, and the first and second values respectively correspond to high and low binary signal levels.

Claim 9

The neuron unit as claimed in claim 4, wherein said first and second memory means respectively include shift registers each having a storage capacity of at least two bits.

Claim 10

The neuron unit as claimed in claim 7, wherein said first and second memory means respectively use a content of each shift register recursively.

Claim 11

The neuron unit as claimed in claim 4, wherein said first input lines, said first memory means, said first gate means and said third gate means form an excitatory group, and said second input lines, said second memory means, said second gate means and said fourth gate means form an inhibitory group.

Claim 12

The neuron unit as claimed in claim 9, wherein said third and fourth gate means and said output means form majority determining means for determining the output signal of the neuron unit based on a majority of outputs obtained from the excitatory group and outputs obtained from the inhibitory group.

Claim 13

The neuron unit as claimed in claim 3, wherein said input line means includes a plurality of input lines for receiving binary input signals which undergo transitions with time, said supplying means includes memory means for storing the weight functions and corresponding grouping information, said grouping information indicating one of excitatory and inhibitory groups to which said weight functions belong, and said operation means includes: first gate means for successively obtaining a logical product of one of said binary input signals received from said input lines and a corresponding one of the weight functions read out from said memory means for each of said binary input signals; second gate means for obtaining a logical product of one of said grouping information read out from said memory means and a corresponding one of logical products output from said first gate means for each of the logical products output from said first gate means; third gate means for obtaining a logical product of an inversion of one of said grouping information read out from said memory means and a corresponding one of the logical products output from said first gate means for each of the logical products output from said first gate means; fourth gate means for obtaining a logical sum of logical products output from said second gate means; fifth gate means for obtaining a logical sum of logical products output from said third gate means; and output means including an inverter for inverting the logical sum output from said fifth gate means and a gate for obtaining one of a logical product and a logical sum of the logical sum output from said fourth gate means and an inverted logical sum output from said inverter, said gate outputting the output signal of the neuron unit.

Claim 14

The neuron unit as claimed in claim 11, wherein said binary input signals respectively describe a signal quantity in a form of a pulse density, and said memory means stores the weight functions which respectively describe a signal quantity in a form of a pulse density.

Claim 15

The neuron unit as claimed in claim 12, wherein each weight function has a pulse density defined by a number of first values and second values within a predetermined time, the first values and the second values are arranged at random, and the first and second values respectively correspond top high and low binary signal levels.

Claim 16

The neuron unit as claimed in claim 11, wherein said memory means includes a storage capacity of at least three bits for one weight function and a corresponding grouping information.

Claim 17

The neuron unit as claimed in claim 14, wherein said memory means includes shift registers each having a storage capacity of at least two bits for storing the weight functions, and 1-bit memories for storing the grouping information.

Claim 18

The neuron unit as claimed in claim 15, wherein said memory means respectively use a content of each shift register recursively.

Claim 19

The neuron unit as claimed in claim 3, wherein said input line means includes a plurality of input lines for receiving binary input signals which undergo transitions with time, said supplying means includes first and second memory means for storing the weight functions, and said operation means includes: first gate means for successively obtaining a logical product of one of said binary input signals received from said input lines and a corresponding one of the weight functions read out from said first memory means for each of said binary input signals; second gate means for successively obtaining a logical product of one of said binary input signals received from said input lines and a corresponding one of the weight functions read out from said second memory means for each of said binary input signals; third gate means for obtaining a logical sum of logical products output from said first gate means; fourth gate means for obtaining a logical sum of logical products output from said second gate means; and output means including an inverter for inverting the logical sum output from said fourth gate means and a gate for obtaining one of a logical product and a logical sum of the logical sum output from said third gate means and an inverted logical sum output from said inverter, said gate outputting an output signal of said neuron unit.

Claim 20

The neuron unit as claimed in claim 17, wherein said binary input signals respectively describe a signal quantity in a form of a pulse density, and second first and second memory means store the weight functions which respectively describe a signal quantity in a form of a pulse density.

Claim 21

The neuron unit as claimed in claim 18, wherein each weight function has a pulse density defined by a number of first values and second values within a predetermined time, the first values and the second values are arranged at random, and the first and second values respectively correspond to high and low binary signal levels.

Claim 22

The neuron unit as claimed in claim 17, wherein said first and second memory means respectively include shift registers each having a storage capacity of at least two bits.

Claim 23

The neuron unit as claimed in claim 20, wherein said first and second memory means respectively use a content of each shift register recursively.

Claim 24

The neuron unit as claimed in claim 17, wherein said first memory means, said first gate means and said third gate means form an excitatory group, and said second memory means, said second gate means and said fourth gate means form an inhibitory group.

Claim 25

The neuron unit as claimed in claim 22, wherein said third and fourth gate means and said output means form majority determining means for determining the output signal of the neuron unit based on a majority of outputs obtained from the excitatory group and outputs obtained from the inhibitory group.

Claim 26

The neuron unit as claimed in claim 3, wherein said generating means of said self-learning means generates the new weight functions based on the errors and a learning constant.

Claim 27

The neuron unit as claimed in claim 24, which further comprises means for arbitrarily setting the learning constant from outside the neuron unit.

Claim 28

The neuron unit as claimed in claim 3, which further comprises switching means for switching a mode of the neuron unit between first and second modes, said varying means of said self-learning means being enabled in the first mode to thereby renew the weight functions, said varying means of said self-learning means being disabled in the second mode to thereby fix the weight functions.

Claim 29

The neuron unit as claimed in claim 3, wherein said supplying means of said forward process means includes memory means for storing the weight functions, and the neuron unit further comprises means for making access to said memory means, so that the weight functions can be written into and read out from said memory means from outside the neuron unit.

Claim 30

The neuron unit as claimed in claim 3, which further comprises first memory means for storing the input signals, second memory means for storing the teaching signals, and control means for controlling said first and second memory means to supply the stored input signals to said forward process means and the teaching signals to said self-learning means.

Claim 31

The neuron unit as claimed in claim 28, wherein said control means includes counter means for counting a number of learning processes carried out by said self-learning means.

Claim 32

The neuron unit as claimed in claim 3, wherein said supplying means of said forward process means includes memory means for storing the weight functions, and the neuron unit further comprises setting means for variably setting a data length of the weight functions in said memory means.

Claim 33

The neuron unit as claimed in claim 30, wherein each of the input signals and the weight functions respectively describe a signal quantity in a form of a pulse density, and said setting means variably sets a length of each pulse train which corresponds to the weight function.

Claim 34

The neuron unit as claimed in claim 31, wherein said memory means includes a shift register for storing the weight function, and said setting means variably sets a bit length to be shifted in said shift register.

Claim 35

The neuron unit as claimed in claim 3, which further comprises first converter means for converting analog input signals into digital input signals which are supplied to said forward process means as the input signals, and second converter means for converting the output signal of the neuron unit into an analog signal.

Claim 36

The neuron unit as claimed in claim 3, which further comprises switching means for switching at least one of the input signals to said forward process means and the output signal of said self-learning means in response to a plurality of external signals.

Claim 37

The neuron unit as claimed in claim 3, wherein said operation means of said forward process means includes first means for successively carrying out an operation on the input signals in groups of the input signals, second means for successively storing results of operations for the groups, and third means for carrying out an operation on the results of operations simultaneously read out from said second means.

Claim 38

The neuron unit as claimed in claim 3, wherein said operation means of said forward process means includes first means for successively carrying out an operation on the input signals in groups of the input signals, second means for successively storing results of operations for the groups, third means for carrying out an operation the input signals in a first mode and for carrying out an operation on the results of operations simultaneously read out from said second means in a second mode, and fourth means for setting a mode of said third means to one of the first and second modes.

Claim 39

The neuron unit of claim 3, wherein: the errors are pulse-density-modulated.

Claim 40

A neural network, comprising: A) a plurality of neuron units which are coupled to form a hierarchical structure which has a plurality of layers, each of the neuron units simultaneously processing a plurality of binary input signals and outputting an output signal which is indicative of a result of the processing, the neuron units including: a) input line means for receiving the input signals, the input signals being binary pulse-density-modulated signals; b) forward process means, coupled to the input line means, and including: 1) supplying means, made up of digital circuits, for supplying weight functions; and 2) operation means, made up of digital circuits, for carrying out an operation on each of the input signals using one of the weight functions and for outputting the output signal, the output signal being a binary pulse-density-modulated signal; and c) self-learning means, coupled to the forward process means, and including: 1) generating means for generating new weight functions based on errors between the output signal of the forward process means and teaching signals; and 2) varying means for varying the weight functions supplied by the supplying means of the forward process means to the new weight functions generated by the generating means; and B) a plurality of signal lines coupling outputs of arbitrary neuron units in one layer of the hierarchical structure to inputs of arbitrary neuron units in another layer of the hierarchical structure.

Claim 41

A neural network, comprising: A) a plurality of neuron units which are coupled to form a hierarchical structure which has a plurality of layers, each of the neuron units simultaneously processing a plurality of binary input signals and outputting an output signal which is indicative of a result of the processing, the neuron units including: a) input line means for receiving the input signals, the input signals being binary pulse-density-modulated signals; b) forward process means, coupled to the input line means, and including: 1) supplying means, made up of digital circuits, for supplying weight functions; and 2) operation means, made up of digital circuits, for carrying out an operation on each of the input signals using one of the weight functions and for outputting the output signal, the output signal being a binary pulse-density-modulated signal; and c) self-learning means, coupled to the forward process means, and including: 1) generating means for generating new weight functions based on errors between the output signal of the forward process means and teaching signals; and 2) varying means for varying the weight functions supplied by the supplying means of the forward process means to the new weight functions generated by the generating means; and B) a plurality of signal lines coupling outputs of arbitrary neuron units in one layer of the hierarchical structure to inputs of arbitrary neuron units in another layer of the hierarchical structure.

Claim 42

The neural network of claim 38, wherein: the errors are pulse-density-modulated.

Claim 43

The neural network as claimed in claim 38, wherein said operation means of said forward process means includes first means for successively carrying out an operation on the input signals in groups of the input signals, second means for successively storing results of operations for the groups, and third means for carrying out an operation on the results of operations simultaneously read out from said second means.

Claim 44

The neural network as claimed in claim 38, wherein said operation means of said forward process means includes first means for successively carrying out an operation on the input signals in groups of the input signals, second means for successively storing results of operations for the groups, third means for carrying out an operation the input signals in a first mode and for carrying out an operation on the results of operations simultaneously read out from said second means in a second mode, and fourth means for setting a mode of said third means to one of the first and second modes.

Claim 45

The neural network as claimed in claim 38, wherein said input line means includes a plurality of first input lines for receiving first binary input signals which undergo transitions with time and a plurality of second input lines for receiving second binary input signals which undergo transitions with time, said supplying means includes first and second memory means for storing the weight functions, and said operation means includes: first gate means for successively obtaining a logical product of one of said first binary input signals received from said first input lines and a corresponding one of the weight functions read out from said first memory means for each of said first binary input signals; second gate means for successively obtaining a logical product of one of said second binary input signals received from said second input lines and a corresponding one of the weight functions read out from said second memory means for each of said second binary input signals; third gate means for obtaining a logical sum of logical products output from said first gate means; fourth gate means for obtaining a logical sum of logical products output from said second gate means; and output means including an inverter for inverting the logical sum output from said fourth gate means and a gate for obtaining one of a logical product and a logical sum of the logical sum output from said third gate means and an inverted logical sum output from said inverter, said gate outputting the output signal of the neuron unit.

Claim 46

The neural network as claimed in claim 42, wherein said first and second binary input signals respectively describe a signal quantity in a form of a pulse density, and said first and second memory means stores the weight functions which respectively describe a signal quantity in a form of a pulse density.

Claim 47

The neural network as claimed in claim 43, wherein each weight function has a pulse density defined by a number of first values and second values within a predetermined time, the first values and the second values are arranged at random, and the first and second values respectively correspond to high and low binary signal levels.

Claim 48

The neural network as claimed in claim 42, wherein said first and second memory means respectively include shift registers each having a storage capacity of at least two bits.

Claim 49

The neural network as claimed in claim 45, wherein said first and second memory means respectively use a content of each shift register recursively.

Claim 50

The neural network as claimed in claim 42, wherein said first input lines, said first memory means, said first gate means and said third gate means form an excitatory group, and said second input lines, said second memory means, said second gate means and said fourth gate means form an inhibitory group.

Claim 51

The neural network as claimed in claim 47, wherein said third and fourth gate means and said output means form majority determining means for determining the output signal of the neuron unit based on a majority of outputs obtained from the excitatory group and outputs obtained from the inhibitory group.

Claim 52

The neural network as claimed in claim 38, wherein said input line means includes a plurality of input lines for receiving binary input signals which undergo transitions with time, said supplying means includes memory means for storing the weight functions and corresponding grouping information, said grouping information indicating one of excitatory and inhibitory groups to which said weight functions belong, and said operation means includes: first gate means for successively obtaining a logical product of one of said binary input signals received from said input lines and a corresponding one of the weight functions read out from said memory means for each of said binary input signals; second gate means for obtaining a logical product of one of said grouping information read out from said memory means and a corresponding one of logical products output from said first gate means for each of the logical products output from said first gate means; third gate means for obtaining a logical product of an inversion of one of said grouping information read out from said memory means and a corresponding one of the logical products output from said first gate means for each of the logical products output from said first gate means; fourth gate means for obtaining a logical sum of logical products output from said second gate means; fifth gate means for obtaining a logical sum of logical products output from said third gate means; and output means including an inverter for inverting the logical sum output from said fifth gate means and a gate for obtaining one of a logical product and a logical sum of the logical sum output from said fourth gate means and an inverted logical sum output from said inverter, said gate outputting the output signal of the neuron unit.

Claim 53

The neural network as claimed in claim 49, wherein said binary input signals respectively describe a signal quantity in a form of a pulse density, and said memory means stores the weight functions which respectively describe a signal quantity in a form of a pulse density.

Claim 54

The neural network as claimed in claim 50, wherein each weight function has a pulse density defined by a number of first values and second values within a predetermined time, the first values and the second values are arranged at random, and the first and second values respectively correspond top high and low binary signal levels.

Claim 55

The neural network as claimed in claim 49, wherein said memory means includes a storage capacity of at least three bits for one weight function and a corresponding grouping information.

Claim 56

The neural network as claimed in claim 52, wherein said memory means includes shift registers each having a storage capacity of at least two bits for storing the weight functions, and 1-bit memories for storing the grouping information.

Claim 57

The neural network as claimed in claim 53, wherein said memory means respectively use a content of each shift register recursively.

Claim 58

The neural network as claimed in claim 38, wherein said input line means includes a plurality of input lines for receiving binary input signals which undergo transitions with time, said supplying means includes first and second memory means for storing the weight functions, and said operation means includes: first gate means for successively obtaining a logical product of one of said binary input signals received from said input lines and a corresponding one of the weight functions read out from said first memory means for each of said binary input signals; second gate means for successively obtaining a logical product of one of said binary input signals received from said input lines and a corresponding one of the weight functions read out from said second memory means for each of said binary input signals; third gate means for obtaining a logical sum of logical products output from said first gate means; fourth gate means for obtaining a logical sum of logical products output from said second gate means; and output means including an inverter for inverting the logical sum output from said fourth gate means and a gate for obtaining one of a logical product and a logical sum of the logical sum output from said third gate means and an inverted logical sum output from said inverter, said gate outputting an output signal of said neuron unit.

Claim 59

The neural network as claimed in claim 55, wherein said binary input signals respectively describe a signal quantity in a form of a pulse density, and second first and second memory means store the weight functions which respectively describe a signal quantity in a form of a pulse density.

Claim 60

The neural network as claimed in claim 56, wherein each weight function has a pulse density defined by a number of first values and second values within a predetermined time, the first values and the second values are arranged at random, and the first and second values respectively correspond to high and low binary signal levels.

Claim 61

The neural network as claimed in claim 55, wherein said first and second memory means respectively include shift registers each having a storage capacity of at least two bits.

Claim 62

The neural network as claimed in claim 58, wherein said first and second memory means respectively use a content of each shift register recursively.

Claim 63

The neural network as claimed in claim 55, wherein said first memory means, said first gate means and said third gate means form an excitatory group, and said second memory means, said second gate means and said fourth gate means form an inhibitory group.

Claim 64

The neural network as claimed in claim 60, wherein said third and fourth gate means and said output means form majority determining means for determining the output signal of the neuron unit based on a majority of outputs obtained from the excitatory group and outputs obtained from the inhibitory group.

Claim 65

The neural network as claimed in claim 38, wherein said generating means of said self-learning means generates the new weight functions based on the errors and a learning constant.

Claim 66

The neural network as claimed in claim 62, wherein said neuron unit further comprises means for arbitrarily setting the learning constant from outside the neuron unit.

Claim 67

The neural network as claimed in claim 38, wherein said neuron unit further comprises switching means for switching a mode of the neuron unit between first and second modes, said varying means of said self-learning means being enabled in the first mode to thereby renew the weight functions, said varying means of said self-learning means being disabled in the second mode to thereby fix the weight functions.

Claim 68

The neural network as claimed in claim 38, wherein said supplying means of said forward process means includes memory means for storing the weight functions, and the neuron unit further comprises means for making access to said memory means, so that the weight functions can be written into and read out from said memory means from outside the neuron unit.

Claim 69

The neural network as claimed in claim 38, wherein said neuron unit further comprises first memory means for storing the input signals, second memory means for storing the teaching signals, and control means for controlling said first and second memory means to supply the stored input signals to said forward process means and the teaching signals to said self-learning means.

Claim 70

The neural network as claimed in claim 66, wherein said control means includes counter means for counting a number of learning processes carried out by said self-learning means.

Claim 71

The neural network as claimed in claim 38, wherein said supplying means of said forward process means includes memory means for storing the weight functions, and the neuron unit further comprises setting means for variably setting a data length of the weight functions in said memory means.

Claim 72

The neural network as claimed in claim 68, wherein each of the input signals and the weight functions respectively describe a signal quantity in a form of a pulse density, and said setting means variably sets a length of each pulse train which corresponds to the weight function.

Claim 73

The neural network as claimed in claim 69, wherein said memory means includes a shift register for storing the weight function, and said setting means variably sets a bit length to be shifted in said shift register.

Claim 74

The neural network as claimed in claim 38, wherein said neuron unit further comprises first converter means for converting analog input signals into digital input signals which are supplied to said forward process means as the input signals, and second converter means for converting the output signal of the neuron unit into an analog signal.

Claim 75

The neural network as claimed in claim 38, wherein said neuron unit further comprises switching means for switching at least one of the input signals to said forward process means and the output signal of said self-learning means in response to a plurality of external signals.