Optical neural network

• US 5,095,459 A
• Filed: 07/05/1989
• Issued: 03/10/1992
• Est. Priority Date: 07/05/1988
• Status: Expired due to Fees

• Current Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Inventors: Ohta, Jun, Oita, Masaya, Kyuma, Kazuo, Tai, Shuichi
• Primary Examiner(s): Smith, Jerry
• Assistant Examiner(s): Trammell, Jim
• Application Number:
  US07/375,813
• Time in Patent Office: 979 Days
• Field of Search: 364/807, 364/822, 364/837, 364/841, 364/842, 364/845, 364/513, 382/11, 382/12, 350/96.11-96.14, 350/162.12-162.15
• US Class Current: 706/40
• CPC Class Codes: G06N 3/067 using optical means
• CPC Class Codes: G06V 10/88 Image or video recognition using optical means, e.g. reference filters, holographic masks, frequency domain filters or spatial domain filters

Abstract

An optical neural network which imitates a biological neural network, to provide an associative and/or pattern recognition function, is made of light emitting elements to represent an input neuron state vector, a correlation matrix which modulates light according to stored vector information, light receiving elements, an accumulator and a comparator to perform a threshold function. A stored vector closest to an input vector can be found from a large amount of information without increasing the system size by dividing the correlation matrix and the input neuron state vector with time division techniques, frequency modulation or phase modulation techniques. Positive and negative valves can also be provided with similar techniques.

Claims

1. An optical neural network for selecting a vector most similar to an input vector from stored vector information by performing a product operation of said input vector and a correlation matrix representing said vector information, said optical neural network comprising:

means for generating an optical signal,

means for modulating said optical signal,

means for receiving said modulated optical signal, and

a control unit having

means for dividing said input vector into a plurality of partial vectors and said correlation matrix into a plurality of two dimensional submatrices, including means for controlling at least one of said means for generating an optical signal, said means for modulating, and said means for receiving, to multiply said partial vectors and said submatrices in time sequence thereby providing a product of said correlation matrix and said input vector.

2. An optical neural network as set forth in claim 1, wherein

said means for generating an optical signal produces a signal indicative of said input vector;

said means for modulating said optical signal modulates said signal with a function of said stored vector information to produce a modulated optical signal indicative of the partial product of said input vector and said stored vector information;

said means for receiving receives said modulated optical signal from said means for modulating, and provides an electrical signal corresponding to said modulated optical signal, and

said optical neural network further comprises;

means for accumulating and receiving an output of said means for receiving and for providing an output signal indicative of the partial product of said input vector and said stored vector information;

a threshold device having means for comparing said output of said accumulator with a threshold value indicative of the range of acceptable outputs, and means for modifying the optical signal generated by said means for generating an optical signal on the basis of said comparison; and

wherein said control unit further comprises means for controlling said means for accumulating.

3. An optical neural network for selecting a vector most similar to an input vector from stored vector information by performing a product operation of said input vector and a correlation matrix representing said stored vector information, said optical neural network comprising:

a plurality of light emitting elements for generating an optical signal,

first means for modulating said optical signal,

means for receiving said modulated optical signal, and

a control unit comprising

means for dividing said input vector into a plurality of partial vectors and said stored vector information into a plurality of two dimensional submatrices, and for providing a plurality of frequencies, each frequency corresponding respectively to each said partial vector, and a plurality of pairs of frequencies, each said pair corresponding respectively to each of said submatrices,

second means for modulating each of said light emitting elements respectively with a function of said provided frequencies and the elements of said partial vectors corresponding to said light emitting element,

third means for modulating said first means for modulating with functions of said provided pairs of frequencies and said submatrices to provide a modulated optical signal indicative of modulated partial products of said input vector and said stored vector information, and

means for demodulating said modulated partial products to retrieve a signal indicative of the product of said input vector and said correlation matrix.

4. An optical neural network as set forth in claim 3, wherein

said light emitting elements generate an optical signal indicative of said input vector;

said first means for modulating modulates said optical signal with a function of said stored vector information;

said means for receiving said modulated optical signal provides an electrical signal corresponding to said modulated optical signal;

said optical neural network further comprising means for accumulating said output of said means for demodulating to calculate and output the product of said input vector and said stored vector information; and

means for comparing the output of said accumulating means with a threshold value indicative of the range of acceptable outputs, and having means for modifying the input vector indicated by said means for generating an optical signal on the basis of said comparison.

5. An optical neural network as set forth in claim 3, wherein said neural network is a multilayer neural network consisting of an input layer, a hidden layer, and an output layer.

6. An optical neural network for performing a product and sum operation on a neuron state vector and a correlation matrix, which comprises:

an array of light emitting elements for generating an optical signal indicative of said neuron state vector;

a spatial light modulator for modulating said optical signal generated by said array of light emitting elements to produce a product optical signal indicative of the product of said neuron state vector and said correlation matrix;

means for modulating said spatial light modulator with different frequencies said frequencies indicative of positive and negative values of stored information;

an array of light receiving elements for receiving said product optical signal from said spatial light modulator and for providing an output signal corresponding to said product optical signal;

a signal processor for demodulating said output signal of said array of light receiving elements to retrieve and output signals equivalent to said positive and negative values;

a differential amplifier for calculating the difference between said signals equivalent to said positive and negative values; and

a comparator for comparing said difference from said differential amplifier to a threshold value indicative of a range of desired output values and means for modifying said optical signal generated by said array of light emitting elements on the basis of said comparison.

7. An optical neural network as set forth in claim 6, wherein said means for modulating comprises a driver for modulating and driving said spatial light modulator simultaneously with angular frequencies ω

₁ and ω

₂ respectively indicative of positive and negative values.

8. An optical neural network as set forth in claim 7, wherein said signal processor comprises a band-pass filter for separating said output signal of said array of light receiving elements into components of angular frequencies ω

₁ and ω

₂ to provide signals corresponding to said positive and negative values.

9. An optical neural network as set forth in claim 15, wherein said means for modulating comprises:

a pulse generator which generates pulses for representing positive and negative values; and

a driver responsive to said pulses for driving by time division multiplexing said spatial light modulator.

10. An optical neural network as set forth in claim 9, wherein said signal processor comprises a sample hold amplifier for sampling and holding said output signal of said array of light receiving elements for each positive and negative value in synchronism with said pulse and for providing signals corresponding to said positive and negative values.

11. A method for selecting from stored vector information a vector most similar to an input vector, said stored vector information forming a correlation matrix, said method comprising the steps of:

(a) dividing said input vector into P partial vectors, each partial vector Vp (p=1, 2, . . . , P) having a length associated therewith,

(b) dividing said correlation matrix into S two-dimensional submatrices, forming Q columns of P submatrices, wherein each said submatrix Tp (p=1, 2, . . . , P) in each said column has a number of rows associated therewith, wherein for all p=1 2, . . . , P, said number of rows associated with each said submatrix Tp and said length associated with each said partial vector Vp are equal,

(c) selecting one of said partial vectors and a multiplicatively corresponding submatrix,

(d) generating an optical signal indicative of said selected partial vector by illuminating an array of light emitting elements,

(e) modulating said optical signal with a matrix mask on the basis of said corresponding submatrix to provide a product optical signal indicative of the product of said selected partial vector and said corresponding submatrix,

(f) receiving said product optical signal by light receiving elements, and

(g) adding said product indicated by said optical signal to any previously calculated partial products.

12. A method for an optical neural network having a plurality of light emitting elements, light modulating means and light receiving means, for selecting from stored vector information a vector most similar to an input vector, said stored vector information forming a correlation matrix, said method comprising the steps of:

(a) dividing said input vector into P partial vectors, each partial vector Vp (p=1, 2, . . . , P) having a number of elements, said number less than a number k,

(b) dividing said stored vector information into S two dimensional submatrices, thus forming Q columns of P submatrices, each submatrix Tpq (p=1, 2, . . . P;

1=1, 2, . . . , Q) having a number of rows r associated therewith, whereby for all p=1, 2, . . . , P, said number of rows associated with each submatrix Tpq and said number of elements of each partial vector Vp are equal,

(c) providing P frequencies ω

_p respectively corresponding to each partial vector Vp,

(d) providing P frequencies Ω

p, and Q frequencies Ω

q, respectively corresponding to each submatrix Tpq,

(e) modulating each said light emitting element with a respective modulation function, said modulation function for a light emitting element calculated on the basis of the corresponding elements of said partial vectors and said frequencies ω

_p corresponding to said partial vectors,

(g) modulating said optical signal with a light modulating means with a function of said submatrices and said corresponding frequencies Ω

p and Ω

q,

(h) receiving said modulated optical signal by light receiving elements to output a modulated product signal,

(i) demodulating said modulated product signal to retrieve and output a product signal indicative of a partial product of said input vector and said stored vector information.

13. An optical neural network for performing a product and sum operation on a neuron state vector and a correlation matrix, which comprises:

an array of light emitting elements for generating an optical signal indicative of said neuron state vector;

a spatial light modulator for modulating said optical signal generated by said array of light emitting elements to produce a product optical signal indicative of the product of said neuron state vector and said correlation matrix;

means for modulating said spatial light modulator with time division multiplexing for indicating positive and negative values of stored information;

an array of light receiving elements for receiving said product optical signal from said spatial light modulator and for providing an output signal corresponding to said product optical signal;

a signal processor for demodulating said output signal of said array of light receiving elements to retrieve and output signals equivalent to said positive and negative values;

a differential amplifier for calculating the difference between said signals equivalent to said positive and negative values; and

a comparator for comparing said difference from said differential amplifier to a threshold value indicative of a range of desired output values and means for modifying said optical signal generated by said array of light emitting elements on the basis of said comparison.

14. An optical neural network for performing a product and sum operation on a neuron state vector and a correlation matrix, which comprises:

an array of light emitting elements for generating an optical signal indicative of said neuron state vector;

a spatial light modulator for modulating said optical signal generated by said array of light emitting elements to produce a product optical signal indicative of the product of said neuron state vector and said correlation matrix;

means for modulating said spatial light modulator with different phases, said phases indicative of positive and negative values of stored information;

an array of light receiving elements for receiving said product optical signal from said spatial light modulator and for providing an output signal corresponding to said product optical signal;

a signal processor for demodulating said output signal of said array of light receiving elements to retrieve and output signals equivalent to said positive and negative values;

a differential amplifier for calculating the difference between said signals equivalent to said positive and negative values; and

a comparator for comparing said difference from said differential amplifier to a threshold value indicative of a range of desired output values and means for modifying said optical signal generated by said array of light emitting elements on the basis of said comparison.