Optical interconnector and highly interconnected, learning neural network incorporating optical interconnector therein

Claim 2

A highly interconnected optical neural network having capability to learn, said network comprising: spatial modulating means responsive to an input signal for modulating light to provide modulated optical signals; detecting means for providing detection signals in response to light incident thereon; interconnector means for providing a distribution of modulated interconnecting light beams to said detecting means when said interconnector means is illuminated by modulated optical signals from said modulating means so as to interconnect said modulating means and said detecting means according to a predetermined interconnection pattern, said interconnector means including projections means adapted to project a distribution of interconnecting beams along paths without changing a parameter of each of said beams relative to the other interconnecting beams projected thereby, and interconnection weighting means, provided in said paths so as to be illuminated by said interconnecting beams before said parameter of any of said beams has changed, of controlling said parameter of each interconnecting beam relative to the other interconnecting beams to assign an interconnection weight to each said beam and thereby provide weighted, modulated interconnecting beams; and means responsive to said detection signals for changing a condition of said interconnection weighting means to thereby modify said interconnection weights assigned by said weighting means.

Claim 3

A network as claimed in claim 1, wherein said interconnection weighting means comprises a photosensitive screen which controls the intensity of each interconnecting beam.

Claim 4

A network as claimed in claim 2, wherein said photosensitive screen comprises a photochronic material.

Claim 5

A network as claimed in claim 3, wherein said photochromic material is a photochromic glass.

Claim 6

A network as claimed in claim 2, wherein said photosensitive screen comprises a photodichroic material

Claim 7

A network as claimed in claim 5, further comprising a source of UV light for illuminating said screen.

Claim 8

A network as claimed in claim 6, further comprising an analyzing polarizer disposed between said interconnector means and said detecting means.

Claim 9

A network as claimed in claim 2, wherein said photodichroic material comprises sodium fluoride doped with about 0.178 percent lithium by weight and said screen has a thickness of about 0.2 mm.

Claim 10

A network as claimed in claim 7, wherein said photodichroic material comprises potassium fluoride.

Claim 11

A network as claimed in claim 2, wherein said projection means comprises a page-oriented holographic means having an array of spatially-localized holograms associated in one-to-one correspondence with cell partitions defined on said screen.

Claim 12

A network as claimed in claim 2, wherein said detecting means comprises photodetecting means for generating signals indicative of the intensity of the weighted, modulated interconnecting beams and means for nonlinearly transforming such generated signals according to a sigmoidal function to provide detection signals.

Claim 13

A network as claimed in claim 2, wherein aid photosensitive screen comprises regions having optical transmittivities, and wherein said means for changing a condition of said interconnection weighting means comprises: error signal generation means for generating error signals as a function of (a) said detection signals, (b) signals indicative of a spatial pattern forming a desired output, and (c) modulated optical signals provided by said modulating means and; optical means responsive to said error signals for selecting said regions and varying the transitivity thereof in response to said error signals.

Claim 14

A network as claimed in claim 12, wherein said generation means comprises a computer.

Claim 15

A network as claimed in claim 13, wherein said means for selecting said regions and varying said transitivity thereof comprises: a spatial light modulator disposed so that weighted, modulated interconnecting beams provided by said interconnector means are incident on said spatial light modulator of said varying means in the same pattern that weighted, modulated interconnecting beams are incident on said detecting means and, a phase conjugate mirror means placed in optical communication with said photosensitive screen by said spatial light modulator of said selecting and varying means.

Claim 16

A network as claimed in claim 14, wherein said phase conjugate mirror means provides amplified return beams.

Claim 17

A highly interconnected optical processing apparatus for simulating a highly interconnected neural network by carrying out the matrix operation steps of: (1) AX=Y (2) NL(Y)=Y' (3) E=F(Y', T, X) (4) Z=E 2 Y (5) Δa i j =ce 2 j X i a i j, said apparatus comprising: input means including a spatial light modulator defining pixels for projecting a pattern of modulated light beams representing input data in the form of an input vector X with each modulated beam representing an element X i of X; a detecting means including a photodetecting means which defines pixels that are associated with said pixels of said spatial light modulator, and a nonlinear transforming means; interconnector means disposed in the optical path between said spatial light modulator and said detecting means for providing a distribution of modulated and weighted interconnecting light beams to interconnect each pixel of said spatial light modulator with each pixel of said detecting means when said interconnector means is illuminated by a pattern of modulated beams from said spatial light modulator, said interconnector means including projection means adapted to transmit modulated light beams from said spatial light modulator as interconnecting beams without relatively changing the intensities thereof, and photosensitive screen means for controlling the relative intensities of interconnecting beams incident thereon from said projection means to encode each such beam to have an interconnection strength according to the element of the matrix A, said photodetector detecting the sum of the intensities of the modulated interconnecting beams incident on each pixel thereof to provide summed intensity values representative of the result of the matrix multiplication (b 1) AX=Y where Y is a vector, and said nonlinear transforming means transforming each of the summed intensity values according to a sigmoidal function to provide detection signals which represent the nonlinear vector transformation (2) NL(Y)=Y' where Y') is a vector; and means responsive to input data representative of the vector X, provided detection signals representative of the vector Y', and data indicative of a transmissive vector T, which vector T represents a desired output pattern for modifying a condition of said photosensitive screen means to thereby change the interconnection weights of the modulated interconnecting beams, said means for modifying a condition of said photosensitive screen including means for generating an error vector signal representative of the result of (3) E=F(Y', T, X) and optical means responsive to the error vector signal and the vector Y for generating modifying signals representative of the result of the matrix multiplication (4) Z=E 2 Y; said photosensitive screen means responding to said modifying signals to modify the intensities of the interconnecting beams according to the result Δa.sub.i j =ce.sup.2.sub.j X.sub.i a.sub.i j.

Claim 18

A highly interconnected optical processing apparatus for simulating a highly interconnected neural network by carrying out the matrix operation steps of: (1) AX=Y (2) NL(Y)=Y' (3) E=F(Y', T, X) (4) Z=E 2 Y (5) Δa i j =ce 2 j X i a i j, said apparatus comprising: input means including a spatial light modulator defining pixels for projecting a pattern of modulated light beams representing input data in the form of an input vector X with each modulated beam representing an element X i of X; a detecting means including a photodetecting means which defines pixels that are associated with said pixels of said spatial light modulator, and a nonlinear transforming means; interconnector means disposed in the optical path between said spatial light modulator and said detecting means for providing a distribution of modulated and weighted interconnecting light beams to interconnect each pixel of said spatial light modulator with each pixel of said detecting means when said interconnector means is illuminated by a pattern of modulated beams from said spatial light modulator, said interconnector means including projection means adapted to transmit modulated light beams from said spatial light modulator as interconnecting beams without relatively changing the intensities thereof, and photosensitive screen means for controlling the relative intensities of interconnecting beams incident thereon from said projection means to encode each such beam to have an interconnection strength according to the element of the matrix A, said photodetector detecting the sum of the intensities of the modulated interconnecting beams incident on each pixel thereof to provide summed intensity values representative of the result of the matrix multiplication (b 1) AX=Y where Y is a vector, and said nonlinear transforming means transforming each of the summed intensity values according to a sigmoidal function to provide detection signals which represent the nonlinear vector transformation (2) NL(Y)=Y' where Y') is a vector; and means responsive to input data representative of the vector X, provided detection signals representative of the vector Y', and data indicative of a transmissive vector T, which vector T represents a desired output pattern for modifying a condition of said photosensitive screen means to thereby change the interconnection weights of the modulated interconnecting beams, said means for modifying a condition of said photosensitive screen including means for generating an error vector signal representative of the result of (3) E=F(Y', T, X) and optical means responsive to the error vector signal and the vector Y for generating modifying signals representative of the result of the matrix multiplication (4) Z=E 2 Y; said photosensitive screen means responding to said modifying signals to modify the intensities of the interconnecting beams according to the result Δa.sub.i j =ce.sup.2.sub.j X.sub.i a.sub.i j.

Claim 19

A processing apparatus as claimed in claim 16, wherein said photosensitive screen means comprises a photochromic glass.

Claim 20

A processing apparatus as claimed in claim 16, wherein said photosensitive screen means comprises a photodichroic screen.

Claim 21

A processing apparatus as claimed in claim 18, further comprising a UV light source adapted to completely expose at least one side of said screen and an analyzing polarizor disposed in an optical path from said interconnector means to said photodetecting means.

Claim 22

A processing apparatus as claimed in claim 16, wherein said projection means comprises holographic means in the form of an array of spatially-localized holograms.

Claim 23

A processing apparatus as claimed in claim 20, wherein said computer calculates the error vector E=F(Y', T, X) according to the algorithm E=c(Y'-T) 1/2 where c is the amplification factor of said phase conjugate mirror.

Claim 24

A processing apparatus as claimed in claim 16, wherein said error signal generating means comprises a computer and said responsive optical means comprises a phase conjugate mirror placed in optical communication with said photosensitive screen by a reflective spatial light modulator.

Claim 25

A highly interconnected neural network having capability to learn, said network comprising a plurality of first stage neural network devices which are each adapted to provide an output, and a second stage neural network device receiving as inputs, the output of each of said first stage network devices, each of said devices including means for receiving optical signals as input; detecting means for providing detection signals; interconnector means for providing a distribution of modulated interconnecting light beams to said detecting means when illuminated by optical signals transmitted from said input means in order to interconnect said input means and said detecting means, said interconnector means comprising projection means adapted to project a distribution of interconnecting beams along paths without changing a parameter of the beams relative to each other, and interconnection weighting means including a photosensitive screen located remote from said projection means and in the paths of the interconnecting beams projected therefrom, for controlling said parameter of each interconnecting beam relative to the other interconnecting beams in order to assign an interconnection weight to each such beam and thereby provide weighted, modulated interconnecting beams; and means responsive to detection signals from said detecting means for changing a condition of said interconnection weighting means to thereby modify said interconnection weights.

Claim 26

A highly interconnected neural network having capability to learn, said network comprising a plurality of first stage neural network devices which are each adapted to provide an output, and a second stage neural network device receiving as inputs, the output of each of said first stage network devices, each of said devices including means for receiving optical signals as input; detecting means for providing detection signals; interconnector means for providing a distribution of modulated interconnecting light beams to said detecting means when illuminated by optical signals transmitted from said input means in order to interconnect said input means and said detecting means, said interconnector means comprising projection means adapted to project a distribution of interconnecting beams along paths without changing a parameter of the beams relative to each other, and interconnection weighting means including a photosensitive screen located remote from said projection means and in the paths of the interconnecting beams projected therefrom, for controlling said parameter of each interconnecting beam relative to the other interconnecting beams in order to assign an interconnection weight to each such beam and thereby provide weighted, modulated interconnecting beams; and means responsive to detection signals from said detecting means for changing a condition of said interconnection weighting means to thereby modify said interconnection weights.