Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
  • G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
  • G06N3/0675—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means using electro-optical, acousto-optical or opto-electronic means
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06E—OPTICAL COMPUTING DEVICES; COMPUTING DEVICES USING OTHER RADIATIONS WITH SIMILAR PROPERTIES
  • G06E3/00—Devices not provided for in group G06E1/00, e.g. for processing analogue or hybrid data
  • G06E3/001—Analogue devices in which mathematical operations are carried out with the aid of optical or electro-optical elements

Definitions

• Optical matrix multiplication unit for an optoelectronic system for forming an artificial neural network
• the invention relates to an optical matrix multiplication unit for an optoelectronic system for forming an artificial neural network, with N input waveguides, M output waveguides and a plurality of matrix multiplication unit cells for signal processing of optical signals from one of the N input waveguides and for transmission of the processed signal in each case into one of the M output waveguides, each of the matrix multiplication unit cells being associated with one of the input waveguides and one of the output waveguides and making a one-to-one association between these two associated waveguides.
• the invention further relates to a corresponding matrix multiplication unit cell for such an optical matrix multiplication unit and a corresponding optoelectronic system for forming an artificial neural network.
• US 2020/0110992 A1 describes an optoelectronic system for forming an artificial neural network with an optical matrix multiplication unit, the N input waveguides, M output waveguides and a plurality of matrix multiplication unit cells for signal processing of optical signals from one of the N input waveguides and for transmission of the respectively processed Signal in one of the M output waveguides.
• Each of these matrix multiplication unit cells is associated with one of the input waveguides and one of the output waveguides and makes a one-to-one association between these two associated waveguides.
• the unit cells of this optical matrix multiplication unit each include a Mach-Zehnder interferometer and two phase shifters.
• the optoelectronic system has a light source unit connected upstream of the optical matrix multiplication unit, a modulator unit connected between the light source unit and the matrix multiplication unit, and a sensor unit connected downstream of the matrix multiplication unit.
• phase information of the light used is used for the multiplication.
• the basis for such a procedure is the use of coherent (laser) light.
• the optical matrix multiplication unit according to the invention for an optoelectronic system for forming an artificial neural network, which has N input waveguides, M output waveguides and a plurality of matrix multiplication unit cells for signal processing of optical signals from one of the N input waveguides and for transmission of the processed signal in each case into one of the M output waveguide, in which each of the matrix multiplication unit cells is assigned to one of the input waveguides and one of the output waveguides and makes a one-to-one assignment between these two assigned waveguides, it is provided that each of the matrix multiplication unit cells for signal processing and signal transmission has a connection between the assigned input waveguide and the associated output waveguide interposed directional coupler with electro-optical modulator for transmission control of the directional coupler up ice.
• the multiplication takes place via the amplitude of a corresponding optical signal and not via its phase or a phase relationship.
• the multiplier of the individual multiplication performed by one of the unit cells corresponds to the amplitude ratio of the optical signal between the corresponding input and output waveguides determined by the modulator settings of the electro-optic modulator of this unit cell.
• Such a multiplication based on an amplitude change results in a higher bandwidth, which makes an optoelectronic system with such an optical matrix multiplication unit more powerful.
• the electro-optical modulator of the respective unit cell is a phase modulator.
• the electro-optic modulator (EOM) is based on changing the refractive index. With this method, the refractive index changes by applying an electric field to the doped material. As a result, the phase position of the light changes, which means that the move light waves. Examples of such a phase modulator are Kerr cell and Pockels cell.
• the respective directional coupler has a Mach-Zehnder interferometer, in which the phase modulator is integrated.
• the Mach-Zehnder interferometer has two signal path arms, in one arm of which the phase modulator is arranged.
• the respective directional coupler also has multimode interference couplers for wave splitting at the input and output of the Mach-Zehnder interferometer.
• the electro-optical modulator of the respective unit cell is an absorption modulator.
• This is also referred to as an electro-absorption modulator (EAM).
• EAM electro-absorption modulator
• the opacity of the optical material used is usually changed as a function of an applied voltage.
• the matrix multiplication unit is designed as a semiconductor-based matrix multiplication unit.
• the most common material for this is silicon.
• the matrix multiplication unit is designed as a matrix multiplication unit based on at least one optically active material.
• Possible materials are, for example, lithium niobate, aluminum nitride or gallium nitride.
• the matrix multiplication unit cell according to the invention for an optical matrix multiplication unit mentioned above is intended for signal processing optical signals of an input waveguide of the optical matrix multiplication unit and for the transmission of the respective processed signal in an output waveguide of the optical matrix multiplication unit has a directional coupler with an integrated electro-optical modulator.
• the matrix multiplication unit is designed as the aforementioned matrix multiplication unit.
• FIG. 1 shows a matrix multiplication unit cell for an optical matrix multiplication unit according to a first preferred embodiment of the invention
• FIG. 2 shows the matrix multiplication unit cell for an optical matrix multiplication unit according to a second preferred embodiment of the invention
• FIG. 3 shows an optoelectronic system for forming an artificial neural network with an optical matrix multiplication unit according to a preferred embodiment of the invention.
• FIG. 1 shows a matrix multiplication unit cell 10 for an optical matrix multiplication unit 12 shown in FIG. 3 in a schematic representation.
• the matrix multiplication unit cell 10 is only called unit cell 10 for short below.
• a section of an input waveguide 14 assigned to the unit cell 10 and a section of an output waveguide 16 assigned to the unit cell 10 are also shown.
• These two input and output waveguides 14, 16 are used for the transmission of optical signals and cross one another in the area of the associated unit cell 10, and thus have a crossing point 18.
• input waveguides 14 are drawn in horizontally and output waveguides 16 are drawn in vertically.
• the unit cell 10 For signal processing of the optical signals of the input waveguide 14 (arrow 20) and for transmission of the respectively processed signal (arrow 22) into the output waveguide 16, the unit cell 10 now has a directional coupler 24 which is connected between the input waveguide 14 and the assigned output waveguide 16 and which is equipped with a electro-optical modulator 26 for transmission control of the directional coupler 24 is provided.
• the directional coupler 24 itself is primarily responsible for signal transmission/signal rerouting from the input waveguide 14 into the output waveguide 16 via a signal path 28 .
• Its electro-optical modulator 26 is responsible for the signal processing that makes up the multiplication.
• the electro-optical modulator 26 has electrical connections 30 . In the example shown in FIG.
• the electro-optical modulator 26 is designed specifically as an absorption modulator 32 in which the opacity of the optical material used changes as a function of a voltage applied to the electrical connections 30 .
• the multiplication in this unit cell 10 results from the optical signal from the input waveguide 14 (arrows 20) and its attenuation by the electro-optic modulator 26 to the processed signal (arrows 22).
• 2 shows a variant of the matrix multiplication unit cell 10 in a schematic representation.
• a section of an input waveguide 14 assigned to the unit cell 10 and a section of an output waveguide 16 assigned to the unit cell 10 are shown. These two input and output waveguides 14, 16 are used to transmit optical signals and cross one another in the area of the associated unit cell 10.
• the unit cell 10 has a directional coupler 24 interposed between the input waveguide 14 and the associated output waveguide 16 for signal processing of the optical signals of the input waveguide 14 (arrow 20) and for transmission of the respectively processed signal (arrow 22) into the output waveguide 16.
• an electro-optical modulator 26 for transmission control of the directional coupler 24.
• the electro-optical modulator 26 is in the form of a phase modulator 34 .
• the directional coupler 24 has a Mach-Zehnder interferometer 36 in which the signal path 28 splits into two signal path arms 38, 40 in an intermediate section.
• the phase modulator 34 is located in one of these signal path arms 38.
• the directional coupler 24 at the input and output of the Mach-Zehnder interferometer 36 has multimode interference couplers 42 (MMI couplers) for wave splitting with respect to the signal path arms 38, 40.
• MMI couplers multimode interference couplers
• FIG. 3 shows an optoelectronic system 44 for forming an artificial neural network in a schematic representation.
• the assemblies of this optoelectronic system 44 are shown in a type of block diagram.
• the individual blocks of this block diagram reflect the functional relationships rather than the spatial structure within the optoelectronic system 44 .
• the control electronics required for the controllable components of the individual assemblies are also not shown.
• the assemblies are (i) the matrix multiplication unit 12, (ii) one of Matrix multiplication unit 12 via N input waveguides 14 upstream light source unit 46 and a matrix multiplication unit 12 via M output waveguides 16 downstream sensor unit 48.
• the corresponding N light sources of the light source unit 46 and M sensors of the sensor unit 48 are not shown explicitly.
• the base of the matrix multiplication unit 12 can be made of semiconductor materials such as silicon.
• the matrix itself consists of passive photonic modules for wave guidance, ie the waveguides 14, 16. These waveguides 14, 16 carry a wide range of wavelengths, particularly in the telecommunications sector.
• the waveguides 14, 16 are arranged in rows and columns. A deterministic transfer of optical power from the row to the column waveguides is achieved via the directional couplers 24 .
• the transmission values of the directional couplers 24 encode the matrix elements for the multiplication, i.e. the matrix multiplication unit cells 10. With full transmission, maximum optical power is transmitted into the column and the largest value for the matrix element is displayed; with minimum transmission, the smallest matrix element is realized. Any values in between can be set by controlling the transmission.
• the transmission is controlled by the electro-optical modulators 26. With these, either the real part or the imaginary part of the refractive index is varied. In the case of the real part, it is a phase modulator 34. This is integrated for modulation into the Mach-Zehnder interferometer 36, which consists of two waveguide arms of equal length (Signal path arms 38, 40) is realized.
• the electro-optical phase modulator 34 is integrated into an arm 38 .
• the optical power is divided equally into the two arms 38, 40 with the aid of MMI couplers 42.
• the phase modulator 34 can be implemented, for example, via charge carrier injection in PIN diodes, or also via thermo-optical components.
• Absorption modulators 32 are used to control the imaginary part. On a silicon platform, for example, germanium-based electro-absorption modulators are suitable. These can be modulated at very high speeds in the GHz range and offer a compact design.
• electro-optically active materials such as lithium niobate, aluminum nitride or gallium nitride.
• Efficient waveguides 14, 16 can be produced from these materials, but also efficient electro-optical modulators 26. These work via the electro-optical effect and only consume optical power in the switching state. However, they offer a less compact design than silicon-based modulators 26. However, the switching speed can be in the high GHz range. In addition, these materials offer very broad optical transparency, so that work can also be carried out in the visible wavelength range.
• the invention allows matrix vector multiplication to be performed optically and controlled electrically.
• electro-optical modulators 26 very high switching speeds can be achieved without material fatigue occurring.
• the matrix multiplication unit 12 can thus be configured as often as desired. On the one hand, this allows larger matrices to be created by reprogramming. On the other hand, however, the unit cells 10, ie the corresponding matrix elements, can also be adjusted over time. This is particularly necessary for the optimization of the calculation, as well as for machine learning.
• the modulators 26 provide a high dynamic range that is beyond the electrical Tension can be precisely controlled. This makes it possible to adjust the matrix elements with high accuracy. This also increases the overall result of the
• Matrix multiplication very precise because the input optical power can be precisely controlled.
• High-precision multiplications are essential for machine learning and have so far only been inadequately realized electronically.
• the multiplication approach using the combined optical-electronic variant allows extremely high calculation rates that cannot be achieved with conventional methods. Due to the reprogrammability, the size of the matrix is not limited, so that the invention can be used for effective scaling. At the same time, optical processes offer very high energy efficiency, so that the central challenges of artificial intelligence can be addressed via the process.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Biophysics (AREA)
• Biomedical Technology (AREA)
• General Physics & Mathematics (AREA)
• Mathematical Physics (AREA)
• Health & Medical Sciences (AREA)
• Life Sciences & Earth Sciences (AREA)
• Artificial Intelligence (AREA)
• General Health & Medical Sciences (AREA)
• Neurology (AREA)
• Optics & Photonics (AREA)
• Computational Linguistics (AREA)
• Data Mining & Analysis (AREA)
• Evolutionary Computation (AREA)
• Nonlinear Science (AREA)
• Molecular Biology (AREA)
• Computing Systems (AREA)
• General Engineering & Computer Science (AREA)
• Software Systems (AREA)
• Optical Modulation, Optical Deflection, Nonlinear Optics, Optical Demodulation, Optical Logic Elements (AREA)
• Optical Integrated Circuits (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=77914329

Family Applications (1)

Country Status (5)

Families Citing this family (2)

Family Cites Families (1)

• 2020
  • 2020-09-15 DE DE102020124034.1A patent/DE102020124034A1/de active Pending
• 2021
  • 2021-09-14 WO PCT/EP2021/075232 patent/WO2022058307A1/de unknown
  • 2021-09-14 CN CN202180067867.7A patent/CN116324813A/zh active Pending
  • 2021-09-14 EP EP21777702.8A patent/EP4214641A1/de active Pending
  • 2021-09-14 US US18/044,650 patent/US20230342596A1/en active Pending

Also Published As

Similar Documents

Legal Events