Artificial Neuron Networks

Abstract

An artificial neuron network and corresponding neuron units are described. The neuron network includes two or more layers of a plurality of artificial neuron units. The layers of artificial neuron units are configured to communicate between themselves via two or more optical waveguide (optical fiber) arrangements. The two or more optical waveguide arrangements are configured with predetermined couplings between the two or more waveguides, thereby providing intercommunication between the neuron units of the two or more layers.

Description

[Technical field]

[0001] The present invention relates to optical computing devices, and more particularly to optical computing devices and device configurations suitable for use in optically integrated artificial neuron networks. [Background technology]

[0002] Optical computing involves the manipulation of visible and infrared light, rather than electrical currents as in electronic computing, to perform computations. Optical computing generally achieves faster computation speeds compared to electronic systems, in part because the manipulation of light pulses occurs faster, allowing for higher bandwidth information transmission. For example, due to the much larger dielectric constant in the microwave range compared to the optical range, electrical current signals propagate at around 10% of the speed of light, thus enabling computational speeds of almost 10 times faster in optical computing. Also, photons, unlike electrons, are not polarized and do not carry a charge, so they consume less power and are less susceptible to crosstalk from nearby electric fields.

[0003] Conventional optical processing systems utilize hybrid electro-optical processing, commonly referred to as optoelectronic processing. In these systems, optical signals are used for data transmission or certain processing operations and are converted to electronic signals for other specific processing operations. Such optoelectronic devices can lose about 30% of their energy in converting electron energy to photons and back. Furthermore, converting optical signals to electronic signals and back slows down data transmission and processing. Therefore, there has been an active research effort into all-optical computing, which can eliminate the need for optical-electrical-optical (OEO) conversion, reducing power consumption and increasing processing speed.

[0004] Another advantageous aspect in the field of optical computing is the implementation of artificial neural networks (ANNs). In general, neural network systems provide processing that can solve problems in a manner comparable to the working of the human brain. An artificial neural network is a computer system inspired by the biological neural networks (BNNs) that essentially constitute the brain. The system "learns" to improve its performance and executes a sequence of commands to complete a desired task. More specifically, an ANN evolves a set of relevant characteristics from the learning material provided to it in order to optimize the processing of inputs related to the selected task. A typical ANN system is based on a set of connected units or nodes called artificial neurons, which are the artificial equivalent of the biological neurons that constitute the brain's BNNs. The inter-node connections, which are the artificial equivalent of the biological synapses, can transmit a signal from one node to another. The artificial neurons that receive a signal are configured to process the signal and then transmit a corresponding signal to the artificial neurons to which they are connected. In general, the artificial neurons are arranged in layers. The various layers can perform different types of transformations on the input and transmit a corresponding output signal. The signal is transmitted from the first (input) layer to the last (output) layer, possibly after traversing different layers several times.

[0005] In previous research, the inventors of the present invention have developed a neural network configuration that relays the propagation and coupling of light between multi-core and multi-mode optical fibers.

[0006] For example, WO 2017 / 033197 to Zalevsky et al. teaches an integrated optical module that includes multiple optically coupled channels for use in an artificial neural network (ANN). According to some embodiments, the integrated optical module includes a multi-core optical fiber whose cores are optically coupled.

[0007] WO2019 / 186548 describes an artificial neuron unit and neural network for processing input light, the artificial neuron unit including a modal mixing unit, such as a multimode optical fiber, configured to receive the input light and apply selective mixing to two or more modes of light components in the input light to provide an output light, and a filtering unit configured to apply a preselected filter to the output light to select one or more modes of the output light, thereby providing an output light of the artificial neuron unit. Summary of the Invention

[0008] As mentioned above, the coupling of light propagating through optical fibers can be used for a variety of processing tasks suitable for use in neural network processing. However, there is a need in the art for a functional all-optical neuronal network configuration that can handle both data transmission and processing using optical manipulation. Generally, in conventional optical networks, optical elements are relatively limited in handling nonlinear / processing operations. These functions are currently performed by electronic processing, the use of high energy laser units, and / or the manipulation of cold atoms, but each of these approaches has its drawbacks.

[0009] The present invention provides a processing solution suitable for implementation in optical artificial neuronal networks. The inventive technique enables the design of a fully operational all-optical neuronal network based on the use of multimode and multicore optical fibers and the free-space propagation properties of light.

[0010] Thus, the present invention provides an artificial neuron network comprising a plurality of artificial neurons formed with optical waveguides. The prior art described above with reference to WO2017 / 033197 and WO2019 / 186548 describes the use of multimode and multicore 1D, 2D and 3D waveguides to provide integrated optical modules and artificial neuron units. The present technology further extends the components of the neuron network architecture to provide controlled coupling, processing operations and training related processes, as described in detail below.

[0011] As described in more detail below, the present technology provides artificial optical neuronal units and networks that allow mixing, gain, and selected additional manipulations to be optically applied to signals within the neuronal units. In general, artificial neuronal networks are trained for selected tasks by adjusting weights of signal portions transmitted through various neurons as well as adjusting selected signal paths through the network. The present technology provides optical configurations that provide selective gain / pumping applied to spatial and / or temporal signal portions and selective signal portion mixing that allows weight adjustments to the network.

[0012] According to one broad aspect, the invention provides an artificial neuronal network including a plurality of two or more layers of artificial neuronal units configured to communicate therebetween via an arrangement of two or more optical waveguides (e.g., optical fibers), the arrangement of two or more optical waveguides configured with predetermined couplings between the two or more waveguides, thereby providing intercommunication between the neuronal units of the two or more layers.

[0013] In this regard, it should be noted that the term "waveguide" as used herein relates not only to one-dimensional waveguides, but also to two-dimensional and three-dimensional waveguides. In this regard, one-dimensional waveguides are typically depicted as planar waveguides that support a single transverse mode (e.g., a single-mode optical fiber in a thin multimode optical fiber). Two- or three-dimensional waveguides may support additional transverse modes in one or two transverse axes, up to a bulk waveguide.

[0014] According to some embodiments, at least one of the two or more optical waveguides may have one or more etched patterns to form one or more grating patterns to selectively enhance coupling of optical signals between the at least one waveguide and at least one other waveguide located in selected proximity to the etched pattern.

[0015] According to some embodiments, at least two of the two or more optical waveguides may be configured with a tapered region to provide increased coupling between the at least two optical waveguides in this region.

[0016] According to some embodiments, the tapered region may further include a dedicated interaction region providing free space interaction between optical signals propagating through respective at least two optical waveguides associated with the tapered region. The dedicated interaction region may be formed by a ferrule element. The ferrule element may include a gain medium material capable of being externally pumped, thereby modulating the power of the optical signal propagating within the ferrule element. Additionally or alternatively, the ferrule element may further include an optical reflecting element at one end thereof, thereby providing backscattered light transmitted to at least one of the at least two optical waveguides associated with the tapered region.

[0017] According to some embodiments, the at least two optical waveguides associated with the tapered region may further include a circulator unit selectively defining at least one input waveguide and at least one output waveguide of the tapered region.

[0018] In general, artificial neuronal networks can be configured as cascaded logic gates, where a neuronal network formed with a topology and arrangement of neuronal units undergoes a series of logic gate processing operations on a set of inputs to provide a selected output.

[0019] According to some embodiments, an artificial neuron network includes one or more artificial neuron units, at least one of which includes a modal mixing unit configured to receive input light in a first wavelength range and apply selective mixing to light components of two or more spatial modes of the input light, the modal mixing unit including a multimode optical fiber having a predetermined gain medium impregnated at least a portion thereof and configured to emit light in a first predetermined wavelength range in response to pump light in a second wavelength range, the modal mixing unit further configured to selectively pump additional energy into one or more spatial modes of the input light propagating therethrough in response to pump light in the second wavelength range and one or more selected spatial modes.

[0020] According to some embodiments, an artificial neuron network includes one or more optical processing units, such optical processing units including an optical gain unit having an input facet and an output facet disposed in an optical path between an input multicore optical fiber and an output multicore optical fiber, the optical gain unit, upon exposure to external illumination, generating a holographic pattern within the optical gain unit, thereby selectively affecting optical transmission between the input and output multicore fibers.

[0021] According to some embodiments, an artificial neuron network includes one or more optical processing units, such optical processing units including at least one optical input port for receiving a first optical signal and at least one additional input port for receiving a second additional input signal, the optical processing units including a first optical fiber section, an interaction node, and a second optical fiber section, the first and second optical fiber sections being composed of optical fibers having selected characteristics and lengths for separating an optical signal passing therethrough into wavelength or spatial frequency components, the interaction node being configured to receive signal components of the first optical signal from the first optical fiber section, interacting the signal components with the second additional input signal to generate multiplied signal components, coupling the multiplied signal components to the second fiber section to transform the multiplied signal components, and providing an output signal indicative of the interaction between the first input signal and the second input signal.

[0022] According to some embodiments, an artificial neuron network includes at least one processing junction including an input port adapted to receive first and second input optical signals, and an optical spatial mixing arrangement configured to receive the first and second input optical signals and apply optical processing to provide output data indicative of a correlation between the first and second input optical signals.

[0023] According to another broad aspect, the invention provides an artificial neuron unit including a modal mixing unit configured to receive input light in a first wavelength range and apply selective mixing to light components of two or more spatial modes of the input light, the modal mixing unit including a multimode optical fiber including a predetermined gain medium at least a portion of which is impregnated with a gain medium and configured to emit light in a first predetermined wavelength range in response to pump light in a second wavelength range, the modal mixing unit further configured to selectively pump additional energy into one or more spatial modes of the input light propagating therethrough in response to pump light in the second wavelength range and one or more selected spatial modes.

[0024] According to some embodiments, the artificial neuron unit further comprises a beam combiner arranged at its input, the beam combiner being arranged such that input light in the first wavelength range and pump light in the second wavelength range are coupled into the multimode optical fiber.

[0025] According to yet another broad aspect, the present invention provides a processing junction (or artificial neuron unit) for use in an artificial neuronal network, the processing junction including an input port adapted to receive first and second input optical signals, and an optical spatial mixing structure that receives the first and second input optical signals and applies optical processing to provide output data indicative of a correlation between the first and second input optical signals.

[0026] According to some embodiments, the first and second input optical signals may be associated with optical signals propagating through corresponding first and second multi-core optical fibers.

[0027] According to some embodiments, the optical spatial mixing structure may include at least one optical reflecting element configured to reflect light of the first and second input optical signals into a common spatial path, thereby enabling optical measurement of spatial correlation between the first and second input optical signals.

[0028] According to some embodiments, the processing junction may further comprise a decoherence unit, upstream of the optical spatial mixing structure, configured to reduce spatial coherence of the first and second input optical signals.

[0029] According to some embodiments, the processing junction may be configured to receive the first and second input optical signals via first and second multi-core optical fibers, and the optical spatial mixing structure may be configured to mix the first and second input optical signals with free space propagation of light.

[0030] According to yet another broad aspect, the present invention provides an optical processing unit (or artificial neuron unit) including an optical gain unit having input and output facets disposed in an optical path between an input multi-core optical fiber and an output multi-core optical fiber, said optical gain unit, when exposed to external illumination, generates a holographic pattern within said optical gain unit, thereby selectively affecting optical transmission between said input and output multi-core fibers.

[0031] According to some embodiments, the holographic pattern in the optical gain unit may be three-dimensional.

[0032] According to some embodiments, the optical processing unit may further include an input optical lens unit and an output optical lens unit respectively disposed in the optical path between an input multi-core optical fiber and an input facet of the optical gain unit, and between an output facet of the optical gain unit and the output multi-core optical fiber.

[0033] According to some embodiments, the input and output multi-core optical fibers may be formed as one-dimensional multi-core optical fibers, and the input optical lens unit is an astigmatic lens configured to cause input light to form a three-dimensional spatial pattern within the optical gain unit.

[0034] According to yet another broad aspect, the invention provides an optical processing unit comprising at least one optical input port for receiving a first optical signal and at least one additional input port for receiving a second additional input signal, the optical processing unit comprising a first optical fiber section, an interaction node, and a second optical fiber section, the first and second optical fiber sections being comprised of optical fiber having selected characteristics and lengths for separating an optical signal passing therethrough into wavelength or spatial frequency components, the interaction mode being configured to receive signal components of the first optical signal from the first optical fiber section, interacting the signal components with the second additional input signal to produce multiplied signal components, coupling the multiplied signal components to the second fiber section to transform the multiplied signal components, and providing an output signal indicative of an interaction between the first and second input signals.

[0035] According to some embodiments, the first and second fiber sections may comprise graded refractive index fiber sections having a refractive index profile and length selected to separate an optical signal passing therethrough into its multiple spatial frequencies, thereby applying a spatial Fourier transform to the input signal.

[0036] According to some embodiments, the interaction node may include an arrangement of multiple optical fiber cores formed of optical fiber carrying a gain material, and the second additional input signal may provide selective pumping to the arrangement of multiple optical fiber cores, thereby causing components of the second additional input signal to selectively interact with spatial components of the first optical signal.

[0037] According to some embodiments, the first and second fiber sections may comprise distributed optical fiber having lengths selected to apply a Fourier transform to an optical signal with respect to its wavelength components, and the interacting node may comprise a time modulator configured to receive data indicative of the second additional input signal and to modulate components of the first optical signal accordingly, thereby interacting frequency components of the first and second input signals. [Brief description of the drawings]

[0038] In order to better understand the subject matter disclosed herein, and to illustrate how it may be carried out in practice, embodiments will now be described, by way of non-limiting example only, with reference to the accompanying drawings, in which: [Figure 1] FIG. 1 shows the general structure of an artificial neuron network. [Diagram 2] FIG. 2 is a diagram illustrating the coupling of optical signals between optical fibers according to some embodiments of the present invention. [Diagram 3] FIG. 3 is a diagram illustrating an artificial neuron unit capable of non-linear processing according to some embodiments of the present invention. [Figure 4] FIG. 4 illustrates an artificial neuron unit configured to determine correlations between input signals, according to some embodiments of the present invention. [Diagram 5] 5A and 5B illustrate optical convolution and transformation units according to some embodiments of the present invention, with FIG. 5A illustrating a spatial convolution unit and FIG. 5B illustrating a 3D spatial and temporal convolution unit. [Figure 6]6A-6C illustrate artificial neuron unit configurations utilizing bulk gain structures according to some embodiments of the present invention, where FIG. 6A shows an example of 3D patterning of bulk gain structures and direct coupling, FIG. 6B illustrates free-space coupling that allows 2D patterning of bulk gain structures, and FIG. 6C illustrates a bulk gain structure and its pattern as an optical switching unit. [Figure 7] 7A and 7B are diagrams illustrating an artificial neuron unit that utilizes a tapered multi-core optical fiber to apply mixing and gain to selected signal portions, according to some embodiments of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0039] Artificial neuron networks refer to a variety of computing architectures and algorithms inspired by biological neuron networks. Referring to Figure 1, a schematic configuration of an artificial neuron network 100 formed by an array of artificial neuron units (neurons) 200 communicating by links 210 between them is illustrated.

[0040] Generally, the neurons 200 are arranged in multiple layers, including an input layer INL arranged and configured to receive input data for processing, and an output layer OL arranged and configured to provide output data after processing. The network may also include one or more intermediate hidden layers, such as HL1 and HL2 illustrated in FIG. 1. The neurons in one or more intermediate layers bridge between the input layer and the output layer and are responsible for data processing.

[0041] As mentioned above, optical processing networks and optical neuron networks can improve processing speeds compared to electronic processing. In general, the realization of optical (photonic) neural networks is based on two main operations: linear mixing between signals associated with two or more artificial neurons (i.e. traveling along two or more optical fibers or fiber cores) and nonlinear effects associated with amplification and mixing that provide processing within the artificial neuron unit.

[0042] Artificial neuron networks can be configured and trained to perform a variety of computational operations. In general, artificial neuron networks such as network 100 can be configured for a particular processing application based on the topology of the network, the number of layers, the arrangement of the layers, and training of the network for a particular application. For example, artificial neuron network 100 can be configured to operate as a cascaded logic gate. Such logic gates allow for determining output data based on multiple different inputs having a selected relationship between them. Additional exemplary applications may include various data processing applied directly to the optical inputs. For example, processing applications such as image recognition are included. It is noted that while the present invention relates to the configuration of optical neural networks and artificial neuron units, the networks and neuron configurations of the present invention can be used for any processing application for which they are suitable.

[0043] According to some embodiments, the present technology provides an artificial optical neuron unit suitable for operating in an artificial optical neuron network. The artificial neuron unit is configured based on one or more optical fibers or waveguides and configured to apply amplification gain, mixing, and / or selected operations / transformations to the optical signals. The present technology eliminates or at least significantly reduces the need for optical-to-electronic conversion of the signals, thus substantially eliminating the need for optical-to-electronic signal conversion. The artificial neuron unit according to the present technology is generally formed using one or more optical fibers configured and used for at least one of transmitting optical signal portions, coupling optical signal portions between artificial neuron units, and applying one or more processing operations to the optical signals, thereby operating one or more artificial neuron units.

[0044] Referring to FIG. 2, an optical link unit 210a is illustrated formed of an array of optical fibers 220a-220n. The optical link unit provides for the transmission of an optical signal to provide an input signal to the artificial neuron unit. The optical link unit 210a is formed in one or more selected optical fiber sections and in close proximity to one or more other optical fiber sections. In this embodiment, the fiber 220b is etched with a periodic pattern 250 that forms a grating region (e.g., a Bragg grating) of a selected period and length (e.g., a long period Bragg grating). The periodic pattern 250 is selected to outcouple a preselected signal portion 260 from the optical fiber 220b. The signal portion 260 outcoupled from the optical fiber 220b is coupled to an adjacent optical fiber (e.g., 220a and / or 220c) and the signal portions are mixed. Additionally or alternatively, the signal portion 260 or a portion thereof may be directed out of the optical fiber structure, forming a loss of a selected portion of the optical signal. The signal portions 260 can be selected based on wavelength and / or spatial mode by selecting parameters of the periodic pattern 250 .

[0045] The periodic pattern 250 may be configured as a long period fiber Bragg grating (LPFBG) applied (e.g., etched) to a portion of one or more selected optical fibers 220b or to a portion of the optical fiber core (in the case of a multi-core fiber bundle). The pattern 250 provides controlled coupling between optical fibers with weights for mixing input signal portions selected by the parameters of the pattern 250. The use of the periodic pattern 250 allows for controlled coupling between optical fibers while reducing geometric requirements such as length and physical distance between adjacent optical fibers.

[0046] In some additional configurations, the optical link unit 210a may utilize a periodic pattern 260 formed as a short period fiber Bragg grating (SPFBG). The period of the SPFBG is selected according to the wavelength of the optical signal to separate signal portions based on the wavelength of the signal passing through the optical fiber 220b. The use of SPFBGs along selected optical fibers allows for the use of multi-spectral optical signals to selectively separate signal portions propagating through the optical fiber based on the wavelength of the optical fiber. Such periodic patterns 250 may be advantageously used with optical fibers doped with gain materials (e.g., Erbium-doped optical fibers (EDFAs)) that are used to selectively amplify signal portions, as described in more detail below. For example, in a gain-doped optical fiber with a pump wavelength of 980 nm and a signal wavelength on the order of about 1550 nm, the use of SPFBGs allows for the separation of pump and signal wavelengths by back reflecting the pump wavelength while allowing the signal wavelength to propagate unchanged.

[0047] In some configurations, the core of one or more optical fibers (220b) can be provided with a selected active material having nonlinear properties. Such an active material provides a selectively variable optical effect of the periodic pattern by varying the refractive index profile along the optical fiber. For example, the core of one or more optical fibers 220b can be formed or doped with a photorefractive material (i.e., a material that exhibits a photorefractive effect), a material with Kerr nonlinearity, or a liquid crystal material that allows selective variation of the material refractive index. The refractive index can be changed in response to interacting light by nonlinear interactions of the active material or by using an external field applied to the fiber 220b. For example, the use of such an active material allows selective variation of the optical properties of the periodic pattern 260 by external illumination or by applying an external electric field, in which case the recording of the LPFBG can be reconfigured or tuned via external illumination, for example by changing the refractive index and thus affecting the wavelength of the signal compared to the periodic pattern 250.

[0048] As shown, the artificial neuron unit 200 may be constructed based on a multimode optical fiber (MMF). Such a multimode optical fiber (MMF) is configured with a first end, a second end, and a selected length and diameter, and is used to mix the spatial modes of the propagating signal while propagating the input signal therein. More specifically, the optical field input to the MMF may be a combination of one or more spatial modes for the MMF structure. Each spatial mode propagates through the MMF with a respective group velocity, changing the mode combination of the output light. Furthermore, propagating through the MMF may cause a specific mixing between the optical components depending on the shape and optical properties of the MMF, shifting the optical energy between the modes. Thus, the MMF provides an exit signal associated with the mixing of the modes of the input optical signal.

[0049] Furthermore, according to some embodiments, the MMF may be doped with a gain material selected to provide gain in a wavelength range related to the wavelength of the optical signal passing therethrough (e.g., erodium-doped MMF). In this regard, FIG. 3 shows a schematic representation of an artificial neuron unit 200 according to some embodiments of the present invention. The artificial neuron unit 200 is based on a multimode fiber (MMF) 10 doped with a selected amount of a gain medium configured to emit light in a first wavelength range in response to pump light in a second (pumping) wavelength range. The MMF 10 has a first end for receiving input light and a second end for providing output light OL. The MMF 10 may generally be configured as an optical fiber having a single wide core supporting multiple spatial optical modes in the first wavelength range. The artificial neuron unit 200 may also include a beam splitter unit BS configured to receive a first input light wavefront WF related to the optical signal and a second pump light PL. The input light wavefront WF may be formed by a particular spatial optical field. The pumping light PL may have a spatial distribution selected to provide pumping energy to one or more selected spatial modes of light in the MMF 10. The artificial neuron unit 200 may also include an input optical structure 20 and an output optical structure 30 configured to couple light to the MMF 10. Furthermore, the artificial neuron unit 100 may include a spatial light modulator 40 disposed in the optical path of the optical output coupling from the MMF 10, which may also be associated with a control unit configured to selectively change the modulation pattern of the spatial light modulator 40. It should be noted that the input and output light of the artificial neuron unit 200 may be from free space propagating light and / or from light propagating between inked optical fibers.

[0050] The artificial neuron unit 200 typically receives an input optical WF signal coupled to the MMF 10 by the input optical structure 20, mixes the spatial modes of the input optical signal WF to some extent while propagating the input light WF through the MMF 10, and provides an output light EL at its second end. Furthermore, once the selected spatial distribution of the pumping light PL is input to the MMF 10, the pumping light is used to selectively amplify the signal portion associated with the spatial mode having a high overlap with the spatial distribution of the pumping light PL. The output light EL may be further selectively modulated by the spatial light modulator 40 according to the selected operation / task of the artificial neuron unit on which the neuron unit was trained to provide an output optical signal OL. In this regard, it is noted that processing techniques using neural-type configurations are usually based on one or more networks of neuron units. These networks preferably undergo a selected training process in which the internal connections and processing parameters are determined. The artificial neuron unit 100 described herein can be used in various network topologies. For simplicity, the artificial neuron unit 100 may be configured as a single processing unit that can perform selected optical operations by mixing the spatial modes of the input optical signal WF and amplifying the selected spatial mode using the pumping light PL.

[0051] MMF 10 is a multimode fiber having a selected length (e.g., millimeters to centimeters; in some embodiments, MMF 10 can be as long as several meters) and diameter (e.g., 30 micrometers or more, 50 micrometers or more), typically configured to support the propagation of light in a selected wavelength range (e.g., 1.5 micrometers) propagating in multiple spatial modes. The core of MMF 10 is doped with a material having gain characteristics at a selected doping ratio, for example to provide an erodium or other rare earth doped MMF 10.

[0052] In general, an input optical signal having certain wavefront WF, amplitude and length characteristics is transmitted to the artificial neuron unit 100. The input optical signal WF is coupled to the MMF 10 by the input optical structure 20 and propagates in the MMF 10 towards its second end. In addition, pump light PL of a selected spatial waveform (mode) is also coupled to the MMF 10 to provide optical pumping of the gain medium embedded in the MMF 10. During propagation through the MMF 10, modes with high spatial correlation of the pump light are amplified by stimulated emission of the gain medium. Furthermore, different spatial modes of the optical signal (corresponding to the spatial shape of the input light wavefront WF projected on the structure of the MMF 10) propagate at different velocities and mixing occurs between them. The length of the MMF 10 is selected according to the desired mixing between the spatial modes as well as the level of amplification provided thereby. Typically, for a relatively short MMF 10 (i.e. short with respect to the group velocity dispersion characteristics of the MMF 10), the exit light EL maintains most of the mode content of the input wavefront WF and the selected mode is amplified according to the spatial wavefront of the pump light PL.

[0053] The exit light EL may be directed to a spatial light modulator 40 that applies a selected spatial modulation to the wavefront to provide an output optical OL signal. The output optical OL signal may then be directed to one or more additional neuronal units 200 associated with additional layers of the network and / or to corresponding detection units.

[0054] The input optical structure 20 may be disposed near the input end of the MMF 10 and configured to couple the input light WF into the MMF 10. In some configurations, the input optical structure may be configured to couple pump light into the MMF 10, which may be done by disposing the optical structure 20 downstream of the beam splitter unit BS with respect to the propagation of light into the MMF 10. Alternatively, additional optical structures (not specifically shown) may be used to couple the pump light into the MMF 10. In general, the input optical structure 20 may include one or more optical elements, such as one or more lenses (e.g., an objective lens unit).

[0055] As mentioned above, in some configurations, the artificial neuron unit 200 may also include an output optical structure 30 arranged downstream of the MMF 10 (between the MMF 10 and the spatial light modulator (SLM) 40 and / or downstream of the SLM 40, as illustrated in FIG. 3). This output optical structure 30 may typically be composed of one or more optical elements, such as lenses. The output optical structure is typically configured to collect the output light OL from the artificial neuron unit, influence the divergence and / or propagation direction of the output light (e.g., provide collimated output light) according to a selected path of the output light to additional neurons, and couple to further optical fibers and / or one or more detection units.

[0056] For example, in a typical communication system, the optical signals used are of a first wavelength range of about 1550 nm. Furthermore, a typical erodium-doped optical fiber may emit light at about 1550 nm corresponding to the first wavelength range in response to pump light of a second wavelength of about 980 nm. In this manner, appropriately shaping the spatial distribution of the pump light PL can provide selective pumping of one or more selected spatial modes of the optical signals passing through the MMF 10.

[0057] It should be noted that the artificial neuron unit 200 described herein may be further patterned with one or more periodic patterns selected to induce coupling out or retroreflection of residual pumping light to reduce interference. Such periodic patterns are described herein above with reference to FIG. 2. Alternatively or additionally, the artificial neuron unit 200 or any optical link unit that collects the output light OL may include a spectral filter selected to allow transmission of wavelengths in a first wavelength range while filtering out pumping light in a second wavelength range. The spectral filter may be configured to absorb or deflect the pumping light to prevent transmission of the pumping light between layers of the neural network.

[0058] Certain processing operations may be related to determining the correlation between two input signals. For example, such processing steps may be part of the training of a neural network. Referring to FIG. 4, an optical processing unit 240 for use in an artificial neuron network for determining the correlation between two input signals (e.g., processed labeled data pieces versus processed unlabeled data pieces) is shown. The optical processing unit 240 may operate as an artificial neuron unit 200 in a neural network, configured to receive first and second input optical signals ILa, ILb via corresponding multi-core optical fiber bundles 12a, 12b with selected processing operations. The optical processing unit 240 is configured to spatially superimpose the first and second optical signals ILa, ILb to provide an output light OL indicative of spatial interference between the signals. The output signal OL may be detected by the photodetector 26 or directed to further processing in further processing layers.

[0059] In the specific example of FIG. 4, the optical processing unit 240 includes a lens unit 20 having a selected focal length f and a beam splitter unit BS arranged to receive the input light IL after passing through the lens unit 20. In general, the first part of the correlator optical processing unit 240 has the spatially coherent light and the Fourier transformation by the lens 20. A rotating diffuser 28 may be arranged in the optical path of the input light to break the phase and convert the spatially coherent light into a spatially incoherent signal distribution. As illustrated, the rotating diffuser 28 may be arranged between the lens unit 20 and the beam splitter BS. The second part of the correlator optical processing unit 240 is a cosine transformation performed with a shear interferometer formed by the beam splitter BS, the corner prism 22, and the reflective surface (mirror) 24. The output light OL is in the form of a correlation peak in the spatial correlation between the two input optical signals ILa, ILb. The output light OL may be detected by a single pixel detector 26 or may be transmitted for further use.

[0060] The lens unit 20 is preferably arranged to provide a Fourier image of the first and second input light signals ILa, ILb onto a rotating diffuser 28 (i.e. with a distance f between the outputs of the optical fibers 12a, 12b). The light distribution IL is made incoherent by the rotating diffuser 28 and directed to a beam splitter unit BS. The beam splitter unit BS is arranged to reflect at least a portion R1 of the input light IL towards a corner prism 22, which superimposes a light pattern around the optical axis of the prism 22 and directs a superimposed light portion R2. Such a superimposed light pattern is directed towards a reflecting surface (mirror) 24 as R3 and is reflected again towards the beam splitter BS as R4, which provides an output light OL, in which the first and second portions ILa, ILb of the input light IL are superimposed on each other and interfere with each other.

[0061] Typically, when the first and second input signals ILa, ILb are coincident, the output light OL provides a high correlation peak output. When the input signals are different, the output light OL has a low read out. The optical processing unit 200 illustrated in FIG. 4 can be used, for example, in the training process of an optical neuron network to determine the correlation between the optical signals being processed, while eliminating or at least greatly reducing the need for converting the optical signals into electronic data.

[0062] Some processing functions may involve determining the frequency components of a signal, or generally determining the Fourier transform of a signal. This may involve processing the temporal or spatial frequency components of a signal separately, and performing certain operations such as convolution of two signals. To this end, the artificial neuron network may utilize one or more artificial neuron units that include a selected length of a graded index optical fiber (GRIN fiber) having a selected refractive index profile and length for influencing the light pattern propagating through the optical fiber section to perform one or more selected transformations on the light pattern. For example, the refractive index profile and length may be selected to perform an optical Fourier transform on at least one of the spatial and temporal dimensions of the optical signal coupled thereto. Performing a selected spatial or temporal transformation on an input optical signal may be used for additional operations that can be performed directly on the optical signal. For example, with reference to Figures 5A and 5B, an optical convolution unit 200 is illustrated that performs spatial convolution (Figure 5A) and spatial and temporal convolution (Figure 5B). In this example, the optical convolution unit 200 utilizes a spatial Fourier transform and an erodium-doped optical fiber (EDFA) to determine a spatial convolution between the input optical signal IL and selected weights of spatial frequencies FW provided by selected gain levels (e.g., selected pumping). It converts signal multiplication into a convolution of the Fourier transformed signals, or vice versa. The optical convolution unit 200 is formed by a multimode graded index (GRIN) optical fiber section 50a followed by an EDFA network 512 configured to receive pumping levels having selected radial distributions FW associated with selected weights (e.g., determined by pumping strengths) applied to different spatial frequencies. The EDFA network 512 may be followed by an additional GRIN fiber section 50b to provide an output optical signal OL.

[0063] In some examples, GRIN fiber sections 50a and 50b may preferably have the following refractive index profile: (Formula 1) TIFF0007649559000001.tif13170 where r is the radial coordinate within the fiber, and n1 and n2 are selected refractive index values ​​for the fiber. Using such a refractive index profile, fiber sections 50a and 50b are of length TIFF0007649559000002.tif13170 During the propagation of an optical signal through fiber section 50a or 50b, the variations in refractive index and phase velocity of the signal portion result in a two-dimensional spatial Fourier transform of the signal waveform.

[0064] Thus, an input signal IL having an input waveform structure may be coupled to the first GRIN fiber section 50a. The input signal propagates through the GRIN fiber section 50a, where the signal portions propagate with corresponding group velocities, such that at the output end of the GRIN fiber 50a, an intermediate resulting signal is the spatial Fourier of the input signal. The intermediate resulting signal is coupled to the EDFA network 512 such that different spatial frequencies are coupled to different radial sections of the network 512. As described above, the EDFA network 512 is configured to provide a selected radial distribution of gain (e.g., using a graded doping level or a selected pumping distribution) and thereby apply selected gain levels to different Fourier components of the input signal providing a multiplied Fourier signal. The multiplied Fourier signal is further coupled to the GRIN fiber section 50b and propagates therein while undergoing an inverse Fourier transform to provide output light OL. The output light is a form of a convolution between the input optical signal and the inverse Fourier transform of the gain distribution in the EDFA network 512.

[0065] In general, EDFA network 512 may include an arrangement of NxN fibers appropriately connected to the output of GRIN fiber section 50a and the input of GRIN fiber section 50b for coupling optical portions between the fiber sections. NxN points are at the output plane of the input GRIN fiber and NxN points are at the input plane of the output GRIN fiber. The selected number and arrangement of the NxN optical fibers of EDFA network 512 is related to the resolution of the Fourier transform and convolution processes.

[0066] A dispersive optical fiber provides different phase velocities depending on the wavelength of light passing through it. Such a dispersive optical fiber can therefore apply a selected transformation to the input optical signal with respect to its time frequency. In general, such a dispersive optical fiber performs a temporal Fresnel transformation on the combined input optical signal. A sufficiently long dispersive fiber (i.e., an optical fiber long enough to satisfy the far-field approximation) affects the optical signal passing therethrough as a Fourier transformation in the time domain. FIG. 5B shows an optical convolution unit 200 configured to provide temporal as well as spatial convolution of a signal. In this regard, the intermediate result signal output from the GRIN fiber section 50a is coupled to a temporal convolution section formed by a long dispersive fiber 60a (length not shown to scale), a temporal modulator 612, and a second long dispersive fiber 60b. This temporal convolution section provides a determination of the convolution between the input signal (the intermediate result signal) and the Fourier of the selected modulation applied by the temporal modulator 612. More specifically, the long dispersive fibers 60a and 60b affect the input signal to provide its Fourier transform (or inverse Fourier transform). The time modulator 612 multiplies the time frequency with a selected modulation time pattern and provides a convolution of the signal after the inverse Fourier transform. The time convolution is between the time variation of the input signal and the inverse Fourier of the signal fed to the time modulator 612.

[0067] It should be noted that such GRIN fiber sections (e.g. 50a, as well as long dispersive fibers such as 60a) may be used to form additional processing and conversion units in addition to the convolution units illustrated herein. More specifically, using half of the setup of FIG. 5A or 5B (i.e., removing the time modulators 512 and 612 and replacing them at the output end) allows for determining the spatial and / or temporal Fourier transform of the selected signal. Such a converted signal may be further used for additional processing operations applied to its selected frequency components. Furthermore, additional processing elements (e.g., selective modal gain processing 200 as illustrated in FIG. 3) may be used as intermediate components of such a conversion and convolution unit 200.

[0068] For example, in some configurations, the convolution unit 200 may be configured to generate a modulated clock signal output OL. More specifically, an input clock signal may be used for the input signal IL, where the EDFA network 512 may apply spatial modulation to the clock signal. Additionally or alternatively, a time modulator 612 may be used to apply a selected temporal or wavelength modulation to encode the clock signal in time or wavelength. Such a clock signal may be processed using the three-dimensional space-time transform / convolution unit 200. A selected wavelength-selective filter may be used to filter out undesired wavelengths from the temporal Fourier components in order to apply further processing operations to the frequency components of the input signal IL.

[0069] In some additional configurations of the present technology, the optical processing unit may include a waveguide pattern selectively written / imprinted into the bulk gain structure. In general, such a selected pumping pattern may be optically applied to the bulk gain structure to provide a selected amplification / light manipulation effect according to the pumping pattern. Such a pumping pattern may provide a writing of a waveguide pattern in the bulk gain structure that allows selective filtering and phase velocity manipulation of an input optical signal. Referring to Figs. 6A-6C, a configuration of an artificial neuron unit 200 is illustrated that utilizes a bulk gain structure 17 with a selectively written pattern HW (e.g., volume holographic writing) to provide selected waveguides and selected pumping structures patterned therethrough. The artificial neuron unit 200 includes an input multi-core optical fiber 70a and an output multi-core optical fiber 70b configured to receive an input optical signal IL and direct the input optical signal to be coupled into the bulk gain structure 17, and to receive an optical signal exiting the bulk gain structure 17 and direct the same to provide an output optical signal OL. In the example of Figure 6A, there is a direct coupling, i.e. optical fibre to bulk waveguide coupling, between the multi-core waveguides 70a and 70b and the gain structure 17. In the example of Figure 6B, optical lenses 20 and 30 are used to couple light between the multi-core waveguides 70a and 70b and the gain structure 17. In the example of Figure 6C, the bulk structure is selectively operable as a switch to selectively direct input light into the multi-core waveguide 70b or the multi-core waveguide 70c to provide first or second output optical signals OL1 and OL2.

[0070] The selected writing pattern HW can provide two-dimensional or three-dimensional writing of the waveguides in the gain structure 17. This can be performed using a high-speed laser system to provide a configurable and controllable architecture, where the waveguides and their gain levels can be selectively altered according to the optical input for writing HW. As illustrated in FIG. 6C, selective writing can also be used for selective routing of optical signals. More specifically, an input signal IL can be directed to an output signal OL1 or OL2 according to the current written HW pattern on the gain structure 17 and the interaction of the signal with the pattern so generated within said gain structure 17.

[0071] In general, the bulk gain structure 17 can be patterned using three-dimensional writing. In some configurations, however, the pattern may be two-dimensional to achieve a simpler pattern configuration. Such a configuration is possible in the example of FIG. 6B, where the multicore fibers 70a and / or 70b are one-dimensional fiber arrays carrying one-dimensional optical signals, and the optical lenses 20 and 30 in this configuration can be astigmatic lenses configured to orient the one-dimensional pattern relatively broadly relative to the bulk gain structure 17 to avoid divergence of the optical signal.

[0072] In yet another exemplary configuration according to the present technology, a tapered multi-core fiber can be used to provide direct and appropriate interaction between optical cores. Referring to Figs. 7A and 7B, a controllable artificial neuron unit 200 utilizing a tapered multi-core fiber coupling with a tapered configuration is illustrated. In general, a tapered optical fiber configuration separates the optical fiber cores away from the tapered region to avoid interaction, and brings the cores together at the tapered region to allow interaction between optical signals transmitted therethrough, for example, by optical coupling. Figs. 7A and 7B show an artificial neuron unit 200 having input and output multi-core optical fibers 80a and 80b configured with tapered regions (82a and 82b in Fig. 7A, 82 in Fig. 7B) to provide optical coupling between the cores of the multi-core optical fibers. In the example of Fig. 7A, an input optical signal propagates through the input multi-core fiber 80a with a particular input waveform, and when the multi-core fiber tapers to reach the ferrule 85, the light propagates through the ferrule 85 in a free space mode and is coupled to the output multi-core fiber 80b. This allows coupling and mixing between optical components propagating in different cores of the multicore fiber 80b. The use of a ferrule 85 after the tapered region 82a provides free-space interaction and increased mixing between spatial modes. Additionally, the ferrule can include or be made of a nonlinear material with selected optical properties, such as a gain medium, optical Kerr effect, selected to realize selective pumping of signal portions or change in mixing characteristics. For example, the ferrule 85 may include an erbium gain medium or other rare earth material, allowing external pumping and control of selected weights of the neural network. In some additional configurations, while using erodium optical fibers for the input multicore fiber 80a and / or the output multicore fiber 80b, the ferrule 85 provides free-space mixing between optical components, whereby the input optical signals are first amplified and then mixed and interacted through the tapered device 82a and the ferrule 85.

[0073] In this regard, the example of Figure 7A shows a direct-path unit, while the example of Figure 7B utilises an optical reflector 87 that reflects the optical signal into one end of a ferrule 85 and a circulator 84 disposed between the input 80a and output 80b multicore fibres and coupling them to an intermediate multicore fibre section 80c.

[0074] Thus, the present technique provides an artificial optical neuron unit configured to operate in an artificial neuron network, which allows selective weighting of input signal portions, mixing and manipulation of optical signals, and allows realization of an all-optical or at least nearly all-optical neuron network in which the weight selection, signal processing and even certain training operations are determined directly on the optical signals and do not require conversion to electronic signals.

Claims

1. an artificial neuron network including a plurality of two or more layers of artificial neuron units, the layers of artificial neuron units being configured to communicate therebetween via two or more optical fiber arrangements; an optical link unit is formed within a portion of one or more selected optical fibers of the two or more optical fibers and in proximity to a portion of one or more other optical fibers, and configured to provide predetermined couplings between the optical fibers, thereby providing intercommunication between the neuronal units of the two or more layers.

2. 2. The artificial neuron network of claim 1, wherein one or more of the selected optical fiber portions are etched with a periodic pattern forming a grating region of a selected period and length, the periodic pattern being selected to outcouple selected signal portions from the optical fiber, and the signal portions outcoupled from the optical fiber are coupled to adjacent optical fibers in a manner that mixes the signal portions, thereby providing controlled coupling of optical signals between the optical fibers.

3. 3. The artificial neuron network of claim 1, wherein one or more of the selected optical fiber segments are etched with a periodic pattern forming a grating region of a selected period and length, configured to direct a signal portion or part of it outside the optical fiber and to form a loss of the selected signal portion.

4. 4. The artificial neuron network of claim 2 or 3, wherein the optical fiber comprises a multicore fiber bundle.

5. 4. The artificial neuron network of claim 2 or 3, wherein the periodic pattern is configured as a long-period fiber Bragg grating (LPFBG).

6. 4. The artificial neuron network of claim 2 or 3, wherein the periodic pattern is configured as a short-period fiber Bragg grating (SPFBG).

7. 7. The artificial neuron network of claim 6, wherein the optical fiber carrying the SPFBG is doped with a gain material used for selectively amplifying portions of the signal.

8. 4. An artificial neuron network according to claim 2 or 3, characterized in that the core of the optical fiber is provided with a selected active material having nonlinear properties, said active material providing the selectively variable optical effect of the periodic pattern.

9. 5. The artificial neuron network of claim 4, wherein at least two of the two or more optical fibers are configured with a tapered region in which increased coupling between at least two of the two or more optical fibers is provided.

10. 10. The artificial neuron network of claim 9, wherein the tapered region further comprises a dedicated interaction region providing free space interaction between optical signals propagating through each of the at least two optical fibers associated with the tapered region.

11. 11. The artificial neuron network of claim 10, wherein the dedicated interaction area is formed by a ferrule element.

12. 12. The artificial neuron network of claim 11, wherein the ferrule element includes a gain medium material capable of being externally pumped, thereby modulating the power of an optical signal propagating within the ferrule element.

13. 13. The artificial neuron network of claim 11 or 12, wherein the ferrule element further comprises an optical reflecting element at one end thereof, thereby providing backscattered light transmitted to at least one of the at least two optical fibers associated with the tapered region.

14. 14. The artificial neuron network of claim 13, wherein at least two optical fibers associated with the tapered region further include a circulator unit selectively defining at least one input waveguide and at least one output waveguide of the tapered region.

15. 15. An artificial neuron network according to claim 1, comprising one or more artificial neuron units, at least one of which comprises a modal mixing unit configured to receive input light in a first wavelength range and to apply selective mixing to light components of two or more spatial modes of the input light, the modal mixing unit comprising a multimode optical fiber having a gain medium impregnated at least a portion thereof and configured to emit light in a first predetermined wavelength range in response to pump light in a second wavelength range, the modal mixing unit further configured to selectively pump additional energy into one or more spatial modes of the input light propagating therethrough in response to pump light in the second wavelength range and one or more selected spatial modes.

16. 16. The artificial neuron network of claim 1, comprising at least one optical processing unit, the optical processing unit including an optical gain unit having an input facet and an output facet arranged in an optical path between an input multi-core optical fiber and an output multi-core optical fiber, the optical gain unit being adapted to generate a holographic pattern within the optical gain unit upon exposure to external illumination, thereby selectively affecting optical transmission between the input multi-core optical fiber and the output multi-core optical fiber.

17. 17. The artificial neuron network of claim 1, comprising one or more optical processing units, the optical processing units including at least one optical input port for receiving a first optical signal and at least one additional input port for receiving a second additional input signal, the optical processing units including a first optical fiber section, an interaction node, and a second optical fiber section, the first and second optical fiber sections being made of optical fiber having selected characteristics and lengths for separating an optical signal passing therethrough into wavelength or spatial frequency components, the interaction node being configured to receive a signal component of the first optical signal from the first optical fiber section, interacting the signal component with the second additional input signal to generate a superimposed signal component, coupling the superimposed signal component to the second optical fiber section to transform the superimposed signal component, and providing an output signal indicative of the interaction between the first optical signal and the second additional input signal.

18. 18. An artificial neuron network according to any one of claims 1 to 17, comprising at least one processing junction, the processing junction comprising an input port adapted to receive first and second input optical signals, and an optical spatial mixing structure configured to receive the first and second input optical signals and to apply optical processing to provide output data indicative of a correlation between the first and second input optical signals.

19. 19. An artificial neuron network according to any one of the preceding claims, characterized in that it is adapted to operate as a cascade logic gate.

20. The artificial neuron network of claim 15, further comprising a beam combiner disposed at its input, the beam combiner being arranged such that input light in the first wavelength range and pump light in the second wavelength range are coupled into the multimode optical fiber.

21. The artificial neuron network of claim 18, wherein the first and second input optical signals are associated with optical signals propagating through corresponding first and second multi-core optical fibers.

22. The artificial neuron network of claim 21, wherein the optical spatial mixing structure includes at least one optical reflective element configured to reflect light of the first and second input optical signals into a common spatial path, thereby enabling optical measurement of spatial correlation between the first and second input optical signals.

23. The artificial neuron network of claim 21 or 22, further comprising a decoherence unit configured to reduce spatial coherence of the first and second input optical signals upstream of the optical spatial mixing structure.

24. An artificial neuron network as described in any one of claims 21 to 23, characterized in that the processing junction is configured to receive the first and second input optical signals via first and second multi-core optical fibers, and the optical spatial mixing structure is configured to mix the first and second input optical signals upon optical free-space propagation.

25. The artificial neuron network of claim 16, wherein the holographic pattern in the optical gain unit is three-dimensional.

26. The artificial neuron network of claim 25, further comprising an input optical lens unit and an output optical lens unit respectively disposed in the optical path between an input multi-core optical fiber and an input facet of the optical gain unit, and between an output facet of the optical gain unit and the output multi-core optical fiber.

27. The artificial neuron network of claim 26, wherein the input and output multi-core optical fibers are formed as one-dimensional multi-core optical fibers, and the input optical lens unit is an astigmatic lens configured to cause input light to form a three-dimensional spatial pattern within the optical gain unit.

28. The artificial neuron network of claim 17, wherein the first and second optical fiber sections comprise graduated index fiber sections having a refractive index profile and length selected to separate an optical signal passing therethrough into its multiple spatial frequencies, thereby applying a spatial Fourier transform to an input signal.

29. The artificial neuron network of claim 28, wherein the interacting node includes an arrangement of multiple optical fiber cores formed of optical fiber carrying a gain material, and the second additional input signal provides selective pumping to the arrangement of multiple optical fiber cores, thereby causing components of the second additional input signal to selectively interact with spatial components of the first optical signal.

30. The artificial neuron network of claim 29, wherein the first and second fiber sections comprise dispersive optical fibers having lengths selected to apply a Fourier transform to an optical signal with respect to its wavelength components, and the interacting node comprises a temporal modulator configured to receive data indicative of the second additional input signal and to modulate components of the first optical signal accordingly, thereby interacting frequency components of the first and second input signals.

Images