Method for providing an artificial neural network

Abstract

The invention relates to a method for providing an artificial neural network (200), wherein the following steps are performed: - providing an optical signal (100) for the network (200) in order to obtain an output content (210) of the network (200) by a processing of the optical signal (100) by the network (200), - using a characteristic of the optical signal (100) specific for a spectral phase and / or a temporal phase of the optical signal (100) in order to provide at least one network component (250) of the network (200) for the processing, wherein the at least one network component (250) comprises at least one neuron (251) and / or a weight (252) of the network (200).

Description

Technical Field

[0001] This invention relates to a method for providing an artificial neural network. Furthermore, this invention relates to a system for providing an artificial neural network. Background Technology

[0002] There are numerous potential applications for artificial neural networks, among which the rapid processing of information is crucial. For example, in autonomous driving, real-time and reliable environmental perception is essential. This can be achieved using sensors such as radar, LiDAR, and cameras to detect the vehicle's environment. Typically, a complete 360°-3D detection of the environment can also be specified, allowing all static and dynamic objects to be detected and classified.

[0003] Cameras, in particular, play a crucial role in redundant and robust environmental detection because this type of sensor can accurately measure angles and be used for environmental classification. However, processing and classifying camera images is computationally intensive and structurally costly. This is especially true for 360°-3D environmental detection, where a large number of individual images must be classified and processed, thus increasing computational costs.

[0004] Traditional high-performance artificial neural networks (known as NNs or KNNs) can classify camera images or data from other sensors at image refresh rates of less than 10 Hz. For many applications, such as real-time, reliable environmental monitoring, this refresh rate is often insufficient, as modern camera systems operate at 30 Hz. Furthermore, the data load increases with the increasing image resolution of the camera.

[0005] The limiting factor is usually the processor speed or GPU speed of modern high-performance computers, which, in itself, may not be sufficient to classify images in full real-time. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to at least partially eliminate the aforementioned disadvantages. In particular, the technical problem to be solved by the present invention is to propose alternative solutions for traditional KNN.

[0007] This technical problem is solved by the method according to the invention and by the system according to the invention. Other features and details of the invention are derived from the specification and drawings. Hereinafter, the features and details described in conjunction with the method according to the invention also apply in conjunction with the system according to the invention, and vice versa; therefore, disclosures of individual aspects of the invention are always mutually referenced or can be mutually referenced.

[0008] The subject of this invention is a method for providing artificial neural networks, particularly dispersive neural networks for providing optics. It is particularly specified that the following steps are preferably performed sequentially in the given order and / or repeatedly:

[0009] - Provide optical signals for the network so that the network's output content can be obtained by processing the optical signals by the network, wherein the processing of the optical signals by the network is preferably achieved by means of a predefined optical filter of the signals.

[0010] - Utilizing characteristics of an optical signal, wherein these characteristics may be specific to the phase (spectral and / or temporal) of the optical signal, to provide at least one network component for the processing, wherein the at least one network component preferably comprises at least one neuron and / or at least one weight of the network, wherein the use of the characteristics is preferably achieved by predefined optical filtering comprising a predefined adaptation or adjustment of the characteristics to provide linear weighting and / or nonlinear functions of the neurons.

[0011] - Optionally: Provide multiple neurons as the at least one network component and use a spectral combiner of the network to combine the output signals of the neurons obtained by using the properties in the spectrum.

[0012] This method enables optical processing of signals, thus offering a significant speed advantage over purely electronic processing. Here, the processing, and especially filtering, can be predefined, meaning that the range of changes in the characteristics used for processing can be fixedly preset during network creation (e.g., within the training phase). This can be constructively achieved, for example, by selecting and / or adjusting the optical elements of the network used for the characteristics and / or filtering the optical signal.

[0013] For example, the use of phase-specific properties (spectral and / or temporal) of an optical signal to provide neurons and / or weights can be achieved by using the dispersion of the optical signal in the material during propagation as neurons and / or weights. At least one neuron of an artificial neural network (KNN) can be designed, for example, as an optical and / or dispersive neuron, and / or at least one weight of a KNN can be designed as an optical and / or dispersive weight. In this way, information (especially data) can be processed at the speed of light.

[0014] Optical signals include, for example, optical input signals that are input into the network. Information to be processed can be presented through the input signals, and the information can be processed by processing the optical input signals in order to obtain output content (e.g., in the form of optical output signals, i.e., the processed optical input signals) as the result of the processing.

[0015] One possibility for providing an optical signal is to use a laser pulse as the optical signal (i.e., an optical signal). Alternative implementations of the optical signal, such as a continuous laser beam, are also conceivable. The processing can be achieved by altering the phase distribution (spectral and / or temporal), as will be exemplarily described below. Here, an optical dispersive element adapted to adjust the phase distribution (spectral and / or temporal) can be used for this purpose. Altering the phase distribution (spectral and / or temporal) allows for the use of phase-specific characteristics of the optical signal relative to its (spectral and / or temporal) phase, such as the phase itself, its derivative, or its Fourier transform, as optical neurons and / or as optical weights.

[0016] Furthermore, within the scope of this invention, it is conceivable to provide optical signals by transmitting optical signals to at least one of the at least one network component, wherein the optical signals may be information carriers, and the information is processed by the network to obtain an evaluation, particularly a classification, of the information as output. Therefore, optical signals can be initially input into the network as input signals. Information, such as sensor data, can be input into the network in this manner. Here, information can also be transmitted directly as input signals (e.g., by transmitting light or lidar rays through optical devices) and / or the information can be electronically read in and thus transmitted as input signals on an optical carrier signal. (LiDAR is an abbreviation for light detection and ranging, and can be used, for example, in vehicles to detect the environment and / or measure distance and speed).

[0017] In this invention, it can be advantageously specified that the use of phase-specific characteristics of an optical signal (spectral and / or temporal) is achieved by nonlinearly and / or linearly altering the phase distribution of the optical signal (spectral and / or temporal) in order to (especially in a predefined manner) change said characteristics, in particular by implementing linear weighting of the nonlinear functions (such as activation functions) and / or weights of neurons. Generally, the use of characteristics of the optical signal is achieved, for example, by using the phase (spectral and / or temporal) or related characteristics. The phase-specific characteristics (spectral and / or temporal) of the optical signal can be, respectively, the spectral phase and / or temporal phase itself, or can be group delay (GD), group delay dispersion (GDD), group velocity dispersion (GVD), or third-order dispersion (TOD) of the derivative of the spectral phase, or higher-order dispersion. To perform the aforementioned processing via a network, and specifically via network components, i.e., by implementing weighting and / or neuronal functions (e.g., activation functions) through weights, the aforementioned properties can be altered (e.g., through the use of dispersive materials). For this purpose, optical dispersive elements, as described below, can be used, for example, which can be structurally combined into a network to provide the processing. During weighting, linear changes in the dispersion and / or spectral and / or temporal phase and / or GD and / or GDD and / or TOD can be implemented, for example. Frequency modulation of the optical signal can also be implemented for weighting if necessary. Furthermore, it is conceivable to perform nonlinear changes to the aforementioned properties, such as spectral phase and / or GD and / or GDD and / or TOD, in order to implement the (nonlinear) function of the neurons. Therefore, the neuronal function can be provided, for example, as an activation function, such as a Herveside function, a Sigmoid function, or a ReLU function.

[0018] Within the scope of this invention, the analysis of the output content and / or the output signal of at least one network component can be defined by analyzing, in particular measuring, the phase-specific characteristics (spectral and / or temporal) of the optical signal with respect to the optical signal. To perform the processing via multiple weights and / or neurons, particularly to combine network components (i.e., preferably couple them together), coherent or discontinuous spectral combinations of the optical (output) signals of multiple neurons can be defined, and the combined (output) signals can be transmitted to subsequent optical neurons in the next layer via optical weights. Finally, to analyze the network's output content, the phase-specific characteristics (spectral and / or temporal) of the optical signal can be measured. At the network's output, a dispersion distribution map can be detected as desired output information, for example, classification information in the case of classification, and the output content can be passed, for example, to a CPU (Central Processing Unit) or GPU (graphics processing unit) for further processing, such as forming an environment model.

[0019] The method according to the present invention offers advantages such as increased computational speed for processing via KNNs and the creation of deep optical neural KNNs. Classification can be achieved at the speed of light. Furthermore, lower optical power may be required to trigger the nonlinear response function compared to alternative solutions, as no multi-photon process is needed or only material properties can be used. Furthermore, loop circuitry can be easily implemented. The entire KNN can be physically fabricated on a semiconductor in the form of an optical neural network (ONN). It is also possible to integrate the ONN on a semiconductor chip in CMOS, SiN-CMOS, Bi-CMOS, or hybrid Bi-CMOS processes on a photonics co-integrated chip.

[0020] Furthermore, within the scope of this invention, it can be specified that the optical signal is provided in the form of an optical input signal, which is specific to the input information and processed by the at least one network component to obtain an optical output signal specific to the output content. Therefore, the output signal may include additional information about the input signal, such as classification information, unlike the input signal.

[0021] Further advantageously, it can be specified that multiple network components are provided, each comprising multiple neurons and / or multiple weights, which are arranged in different layers of the network and optically interconnected. The network can typically consist of multiple network components interconnected according to the network structure. Here, the network components can be arranged in different layers of the network, and the output signal of a network component in one layer can be transmitted as an input signal to a network component in a subsequent layer via connections, such as waveguides.

[0022] Optionally, it can be specified that the phase-specific characteristics of the optical signal (spectral and / or temporal) are either the spectral phase and / or temporal phase itself, or group delay, or group delay dispersion, or group velocity dispersion, or third-order dispersion or higher-order dispersion of the derivative of the spectral phase. This enables reliable processing and analysis of the optical signal. Furthermore, it may be feasible to design the neuron as a dispersive neuron, which provides a nonlinear function of the neuron through modulation of the spectral phase of the optical signal.

[0023] According to other advantages, the network can be used in the vehicle, wherein the optical signal is preferably provided in the form of an optical input signal, which is specific to input information about the vehicle's environment and processed by the at least one network component to preferably obtain output content as a classification of the input information. The vehicle may be designed, for example, as a motor vehicle and / or a passenger car (or truck) and / or an autonomous vehicle. The input information may be, for example, signals from sensor data such as those from cameras.

[0024] Similarly, the subject of this invention is a system for providing artificial neural networks, preferably for use in vehicles, the system having:

[0025] - An interface (especially an optoelectronic one) used to provide optical signals for a network, so that the network's output content can be obtained through the network's processing of the optical signals.

[0026] - At least one optical element and / or dispersive element, which is used to utilize the spectral phase and / or time phase-specific characteristics of the optical signal to provide at least one network component of the network for the processing.

[0027] As specified herein, the at least one network component includes at least one neuron and / or weight of the network. Therefore, the method according to the invention can provide a KNN in an optical, dispersive manner.

[0028] Therefore, the system according to the invention provides the same advantages as the method already described in detail with respect to the invention. Furthermore, the system can be adapted to perform the method according to the invention.

[0029] In the system according to the invention, multiple elements can be provided and said elements are structurally coupled to form a network. For this purpose, these elements are integrated into a semiconductor and / or coupled via waveguides, for example, to transmit optical signals between network components. For example, dispersive elements can have a dispersive medium in order to use, in particular, to modify the characteristics of the optical signal.

[0030] Furthermore, it may be advantageous within the scope of the invention to provide a plurality of optical elements, each of which is designed to, in order to utilize the aforementioned characteristics, at least nonlinearly (or alternatively or additionally linearly) change the spectral and / or temporal phase distribution of the optical signal to provide output signals of the neurons of the network (alternatively or additionally, to provide weights), wherein preferably at least one spectral combiner of the network is provided to combine the output signals of the neurons (or weights) spectrally, and preferably a spectral phase analyzer is provided to analyze the phase of the combined output signals in order to provide output content. Attached Figure Description

[0031] Further advantages, features, and details of the invention will become apparent from the following description, in which embodiments of the invention are described in detail with reference to the accompanying drawings. Here, the features mentioned in the claims and description may be important to the invention individually or in any combination. In the drawings, respectively:

[0032] Figure 1 The method steps according to the present invention are shown.

[0033] Figure 2 The neural network is shown.

[0034] Figure 3 A graph showing the Sigmoid function used as a nonlinear function of neurons is shown.

[0035] Figure 4 The electric field of the pulse is shown as a function of time.

[0036] Figure 5 shows the electric field intensity of different laser pulses as a function of time.

[0037] Figure 6 shows the spectral phase of the optical signal as a function of time.

[0038] Figure 7 shows GD and spectral phase as a function of time.

[0039] Figure 8 shows the dispersive element used as a neuron.

[0040] Figure 9 shows a visualization of the propagation of the laser pulse through the quartz glass and its reflection by the dispersive mirror, which acts as a dispersive element.

[0041] Figure 10 shows a visualization of the GD and GDD of the laser pulse after propagation through the quartz glass and the reflection through the dispersive dielectric thin film as a dispersive element.

[0042] Figure 11 shows a distribution of the spectrum and spectral phase of an optical signal processed by a network using a dispersive element.

[0043] Figure 12 shows multiple distributions of the spectrum and spectral phase of the optical signal when processed by a network using a dispersive element.

[0044] Figure 13 An exemplary design of a dispersive optical network is shown.

[0045] Figure 14 An exemplary design of a spectral phase analyzer is shown, in which two delayed pulses are mixed with a time-stretched pulse during a nonlinear process, thereby causing the frequency-converted output pulse to be spectrally clipped.

[0046] Figure 15 Output a flowchart of data processing using optical networks.

[0047] Figure 16 shows the propagation of the laser pulse through the quartz glass and its reflection through a dispersive mirror.

[0048] Figure 17 shows the propagation of the laser pulse through the quartz glass and its reflection through different dispersive mirrors.

[0049] Figure 18 An exemplary design of a neural network formed from semiconductors is shown.

[0050] In the following figures, the same reference numerals are used even for the same technical features in different embodiments. Detailed Implementation

[0051] exist Figure 1The method steps of a method for providing an artificial neural network 200 according to the present invention are schematically visualized. Here, according to a first method step 301, an optical signal 100 (shown as an input signal 101 and an output signal 102 for illustration) is provided to the network 200 so that the output content 210 of the network 200 is obtained by processing the optical signal 100 by the network 200. The input signal 101 may be provided, for example, based on sensor data 103, such as a camera image. Subsequently, according to a second method step 302, characteristics of the optical signal 100, wherein the characteristics are specific to the spectral phase and / or temporal phase of the optical signal 100, are used to provide at least one network component 250 of the network 200 for the processing. As will be described in detail below, the at least one network component 250 may include neurons 251 and / or weights 252 of the network 200.

[0052] As in Figure 1 As further explained, KNN 200 can be used, for example, in the autonomous driving function of vehicle 1 to classify the environment 2 of vehicle 1. Generally, it can be specified that network 200 is used in vehicle 1, wherein optical signals 100 are provided in the form of optical input signals 101, which are specific to the input information regarding the environment 2 of vehicle 1, and are processed by at least one network component 250 to obtain output content 210 classified as the input information. The input information can be determined, for example, by a camera of vehicle 1. Here, sensor data 103, such as camera images from a camera, can be converted into optical input signals 101 and passed to neurons 251 of KNN 200 via weights 252. Figure 2 (Illustrated schematically). During the processing via KNN 200, categories are assigned to individual pixels of the camera image, for example (e.g., to distinguish between roads and road signs, vehicles and pedestrians, or vegetation can also be unique categories). These categories can be represented by output content 210. Through this classification, the environment 2 can be accurately detected, and the output content 210 thus helps in understanding the scene, allowing the driving function to adapt accordingly.

[0053] Here, the output signal of neuron 251, and thus the signal passed to neuron 251 in the next layer, can be given by the sigmoid function, which is the sum of weighted response functions.

[0054]

[0055] in, Representing a weight of 252, σ represents neuron 251, and σ represents the Sigmoid function (see [link to Sigmoid function]). Figure 3 ,exist Figure 3 The image shows an example diagram of the Sigmoid function, which is commonly used as a non-linear function for neurons within a KNN 200. Therefore, the KNN 200 forms a function...

[0056]

[0057] in , and among them, the function value Category information that can be output as output content 210 can be output.

[0058] Traditionally, artificial neural networks are implemented on conventional computer architectures; however, their drawback is slow processing of large amounts of data. In contrast, the optical design of KNN 200 can achieve a significant speed advantage, making it possible to use KNN 200 for driving functions, such as autonomous driving.

[0059] As an optical signal 100, for example, a continuous laser beam (i.e., a continuous wave, i.e., an emission wave that is constant in time) or a laser pulse can be considered, which can be transmitted to at least one network component 250 in a technically reliable manner via waveguide 283. Figure 4 A schematic diagram of a single laser pulse as a function of time t is shown, where the pulse's electric field E oscillates under the envelope curve. Here, a single pulse is characterized by its duration τ, wavelength λ / frequency ω, and amplitude. The minimum achievable duration of such a pulse is defined by the product of time and bandwidth:

[0060]

[0061] As can be seen from equation (3), the pulse has a spectral bandwidth and is therefore a superposition of monochromatic waves of different frequencies. Therefore, as shown in Figure (a) of Figure 5, it is not just a single frequency oscillating below the envelope, but multiple spectral modes oscillating below the envelope.

[0062] The velocity of the envelope is called the group velocity. (Or simply GV) and defined by the derivative of wave number k (wave vector):

[0063]

[0064] The refractive index is given here by n(ω). The propagation speed of each individual monochromatic wave is called the phase velocity. :

[0065]

[0066] Among them, wavenumber through

[0067]

[0068] Provided.

[0069] If the pulse propagates without dispersion, then and The same applies because ∂n(ω) / ∂ω=0 holds (see Figure 5, based on Figure (a) shown in Figure 5). However, this is not the case in dispersive media, and therefore the distribution of the envelope changes (see Figure 5, based on Figure (b) shown in Figure 5). As can be seen from equations (4) and (5), in normally dispersive media (n>1), the delay of the red spectral component is not as strong as that of the blue spectral component, thus stretching the pulse in time (frequency chirp).

[0070] By solving the Helmholtz equations, the electric field E(t) of the laser pulse can be described as a function of time t by the following equations.

[0071]

[0072] in, The carrier frequency is described by ψ(t), the time phase by ψ(t), and the intensity by I(t). For example, for a Gaussian pulse, I(t) is expressed as follows:

[0073]

[0074] The pulse duration is given, where the amplitude E0 of the electric field and the half-value width τ of the pulse define the pulse duration. Equations (7) and (8) show that amplitude, frequency, and pulse duration alone are insufficient to fully characterize the pulse; the time phase must also be taken into account. Spectroscopic observation of the pulse is useful for gaining a deeper understanding of the dispersive dynamics within the pulse.

[0075] By using the Fourier transform of E(t) in the frequency space, equation (7) can be obtained by centering the pulse around its center frequency. get:

[0076]

[0077] Where S( ) represents spectral power density and Representing the spectral phase, the spectral phase can be...

[0078]

[0079] It is indicated that spectral phase defines the phase relationship of individual monochromatic waves below the envelope.

[0080] In Figure 6, different frequencies ω of oscillation are schematically shown in view (a). iThe frequencies oscillate in phase along the envelope of the pulse. According to view (b), the maximum values ​​of each individual frequency are phase-shifted. The spectral phase as a function of time t is shown respectively. In the case of linear phase relationship, all spectral components are in phase, thus constructive interference occurs, and the pulse forms a sharp maximum at t=0. Conversely, linear phase relationship... Corresponding to the time offset, no maximum intensity is formed at t=0 (see Figure 6(b)).

[0081] Equations (9) and (10) clearly show that the dispersion by refractive index has a significant effect on the spectral phase and therefore a significant effect on the intensity distribution of the pulse over time. However, the pulse can be fully characterized by measurements of the spectrum and spectral phase. In this case, the derivative of the spectral phase, such as the group delay (GD):

[0082]

[0083] Group delay dispersion (GDD):

[0084]

[0085] Third-order dispersion (TOD):

[0086]

[0087] Similar quantities have a significant impact on pulse dynamics and are detectable in measurement techniques to describe the interaction between laser pulses and matter. During the propagation of a laser pulse through matter, dispersion can be described by the accumulation of spectral phases. Therefore, short pulses can be temporally chirped by accumulating spectral phase distribution maps. Conversely, chirped pulses can be temporally compressed by accumulating negative phase contributions.

[0088] The following exemplarily describes the spectral phase The possibility of using neuron 251 as a KNN 200. For artificial neuron 251, a non-linear response function is essential. Here, the Sigmoid function is often used in traditional KNNs (see...). Figure 3 The spectral phase distribution of optical signals, such as laser pulses, can be modulated in the same way to map a nonlinear phase distribution. Figure 7 schematically and exemplaryly shows the distribution of the GD in the form of a delta function according to view (a). By integrating the GD, the spectral phase operates in the form of a Herveside function, which may be desirable for use as an optical neuron 251 (see Figure 8, according to view (b)).

[0089] The effects of light-matter interaction in the spectrum, such as the propagation of laser pulses in a dispersive medium, provide the possibility of nonlinearly modulating the spectral phase of incident light waves, similar to the nonlinear modulation of currents via artificial neurons. Here, the dispersive material can be used as an optical neuron 251, which outputs nonlinear phase information. Figure 8 illustrates the operation of the dispersive neuron 251. The incident light signal has a flat phase distribution (see Figure 8(a)). During interaction with the dispersive element, the pulse accumulates spectral phase (see Figure 8(b)). The resulting output spectrum is the same as the input spectrum, but the spectral phase has a nonlinear distribution (see Figure 8(c)).

[0090] Figure 9 shows a simulation of the propagation of a laser pulse before and after passing through a 3 mm quartz glass layer, and its subsequent interaction with a dielectric dispersive layer (such as a chirped mirror). The intensity distribution over time changes due to the light-matter interaction (see Figure 9(a)). However, the spectral distribution of GD has the expected distribution of the delta function, which corresponds to the Heaviside distribution of the spectral phase (see Figure 7). Therefore, by reading out the spectral phase, this information can be used as a neuron 251. Alternatively, GD, GDD, TOD, or higher-order dispersives can be used directly as neurons. The pulse distribution over time t is shown in Figure 9(a), the GD distribution over wavelength λ is shown in Figure 9(b), and the GDD distribution over wavelength λ is shown in Figure 9(c).

[0091] Figure 10 shows the simulation of the Gaussian Dispersion (GD) of a laser pulse after propagation through a 3 mm quartz glass layer according to views (a) and (c), and the simulation of the Gaussian Dispersion (GDD) of a laser pulse after propagation through a 3 mm quartz glass layer according to views (b) and (d), showing the subsequent reflection through the dispersive dielectric layer. The GD distribution pattern exhibits slight oscillations as a function of wavelength. These oscillations can be reduced by optimizing the dispersive dielectric layer. Inserting a special dielectric layer allows the pulse to accumulate more spectral phase, thus linearly reducing the GD to -150 fs in the spectral range of 700 nm to 1100 nm. Therefore, the interaction with these layers is equivalent to subtracting the absolute value of the GD and corresponds to a linear change in the spectral phase. Therefore, operations such as addition, subtraction, multiplication, and division can be performed using the dispersive elements 230, and these elements 230 are implemented as weights 252 in a KNN.

[0092] Figure 11 schematically illustrates the linear modulation of the spectral phase. A spectral filter is used to reduce the spectral bandwidth of the input signal. The resulting spectrum has a smaller bandwidth. This modulation is equivalent to a subtraction operation and affects the subsequent interaction with the dispersive neuron 251, thus preventing the triggering of a nonlinear response in neuron 251 and resulting in a linear phase distribution at the output. Here, view (a) in Figure 11 shows the input spectrum interacting with the spectral filter according to view (b), where the spectral bandwidth is changed, but the phase distribution of the spectrum is not altered (see view (c)). After interaction with the dispersive element 230 (see view (d)), the spectral phase is linearly modulated, which is equivalent to weighting. View (e) shows the resulting spectrum or the resulting phase.

[0093] Figure 12 similarly illustrates optical weighting achieved through frequency shifting, as shown in Figure 11. The spectral phase is multiplied by a constant coefficient through a spectral shift of the input spectrum and subsequent interaction with the dispersive element 230.

[0094] To form the KNN 200, the optical signals 100 of each individual neuron 251 must be transmitted to all neurons 251 in the next layer via weighting 252. An optical spectral combiner 284, particularly a phase combiner 284, provides the possibility for spectral combination of the output signals of the neurons 251. For example, gratings and / or prism sequences and / or dielectric layers and / or optical nonlinear media and / or polarizing optical devices can be used as such components. Here, the spectral phases can be combined coherently or incoherently. Figure 13 The spectral combination of two neurons 251 is illustrated schematically. Individual optical neurons a n m The output signal is spectrally combined using optical components. As the output signal of the spectral combiner, the spectral phase distribution exists as a function of the frequencies of the two individual neurons 251 and can be passed to the next neuron 251 via weights 252. Thus, a dispersive optical neural network (dONN) is formed.

[0095] To perform classification, it may be necessary to measure the spectral phase or its derivative. Various established methods are available for this purpose.

[0096] For example, the interference signal of two optical pulses can be measured using a spectrometer via spectral interferometry, where one of the pulses is delayed by time τ, and the spectral phase of the pulses is known. The combined signal of the two pulses on the spectrometer can be obtained through...

[0097]

[0098] Description. For the oscillating terms, the phase relationship is given by the following equation.

[0099]

[0100] The spectral phase can be reconstructed from the spectral interferogram using the known spectral phase of the reference pulse, and thus the spectrometer can be used as the spectral phase analyzer 285.

[0101] According to another possibility, the heterodyne detector used to characterize the spectral phase can be used as a spectral phase analyzer 285.

[0102] Another possibility lies in using spectral shearing via spectral phase analyzer 285. Here, two delayed pulses are superimposed with a chirped replica pulse in the nonlinear crystal 260 of spectral phase analyzer 285 (see [link to spectral phase analyzer 285]). Figure 14 The frequency conversion in crystal 260 causes spectral shearing of the delayed pulse. GD can be directly extracted from the spectral interferogram.

[0103]

[0104] This includes a spectral shear Ω, which is proportional to the pulse time delay τ. The spectral phase is given by the following equation:

[0105]

[0106] In the FROG structure of the spectral phase analyzer 285, two pulses delayed relative to each other can be superimposed in the nonlinear crystal 260. The newly generated frequencies are recorded by the spectrometer of the spectral phase analyzer 285. The interferogram is obtained as a function of the frequencies and time delays of the two fundamental pulses, and the spectral phase can be reconstructed.

[0107] A complete dONN can be implemented, for example, by connecting at least one of the elements 230 described above. This is exemplarily shown in... Figure 15 Visually displayed. Here, the dispersive element 230 is advantageously considered as the at least one element 230 for forming the optical neuron 251 by spectral phase modulation. Furthermore, as the at least one element 230 for forming the optical weight 252, an element 230 for providing linear modulation of the spectral phase can be used. Additionally, the at least one element 230 may include a spectral combiner 284 designed to coherently or incoherently combine the signals from the dispersive neuron. Furthermore, the at least one element 230 may include a spectral phase analyzer 285, i.e., an element 230 for measuring the spectral phase distribution map. Figure 15 The diagram further shows that electro-optical data processing is implemented by interface 270 to convert sensor data 103 into optical signals 100. Here, sensor data 103 can be transmitted, for example, on an optical carrier (in...). Figure 18 (As shown) Waveguide 283 transmits to network component 250 at different layers 290.

[0108] The following describes other embodiments of the optical element 230 used to provide the network component 250. Thus, a time phase can be applied to the pulse using an electro-optic modulator, such as a Mach-Zehnder modulator (MZM), which serves as element 230. This corresponds to the spectral phase in the frequency space. Therefore, the optical weights 252 can be synthesized by means of linear modulation of the MZM, and the optical neurons 251 can be synthesized by nonlinear modulation of the MZM.

[0109] In addition, a so-called pulse shaper can be used as element 230, which can apply time phase based on LCD or electronic refractive index changes.

[0110] Another possibility for providing network component 250 is to use GD as the response function of the neuron (see Figure 16, where GD before and after propagation is shown according to view (a), and GDD before and after propagation is shown according to view (b), and spectral power density is shown in view (c)) and weights (see Figure 17, where intensity distribution over time is shown according to view (a), and GD before and after propagation is shown according to view (b)).

[0111] Alternatively, GDD can be used as a nonlinear function of neurons 251 and weights 252, or a higher-order dispersion, such as TOD, can be used to provide network component 250.

[0112] Another possibility for providing network component 250 is to focus a high-intensity, low-bandwidth pulse onto the optical nonlinear medium serving as element 230. By using self-phase modulation (SPM), the spectral bandwidth of the pulse can be increased and the pulse can be transmitted to the optical neuron 251. Due to the increased bandwidth, nonlinear modulation of the spectral phase is achieved. Here, SPM is used, for example, as an optical weight. Conversely, a spectral filter can prevent the neuron from being triggered.

[0113] Furthermore, to provide at least one network component 250 or network 200, it is possible to integrate the aforementioned component on semiconductor 280, particularly on an electro-photonic co-integrated chip, using CMOS, bi-CMOS, bybrid bi-CMOS, Si-N CMOS processes, etc. Figure 18As exemplarily shown, optical signal 100 can be optically coupled to network 200 via interface 270. Furthermore, a dispersive layer 281 can be provided, in which network components 250 are disposed. For example, a linear frequency modulator 282 can be used for weighting 252. Additionally, a phase combiner 284 can be used to combine network components 250 from different layers 281. Then, a spectral phase analyzer 285 can provide output content 210 based on the output signals of the network components 250.

[0114] The above description of the embodiments is only an example within the scope of the invention. Of course, if technically useful, the individual features of the embodiments can be freely combined with each other without departing from the scope of the invention.

[0115] List of reference numerals

[0116] 1. Transportation

[0117] 2 Environment

[0118] 100 optical signals

[0119] 101 Input Signal

[0120] 102 Output Signal

[0121] 103 Sensor Data

[0122] 200 Artificial Neural Networks, KNN

[0123] 210 Output content

[0124] 230 components

[0125] 250 network components

[0126] 251 neurons

[0127] 252 weights, weighted average

[0128] 260 crystals

[0129] 270 interface

[0130] 280 Semiconductors

[0131] 281 Dispersion layer

[0132] 282 Linear Frequency Modulator

[0133] 283 Waveguide

[0134] 284 Spectral combiner, phase combiner, grid

[0135] 285 Spectral Phase Analyzer

[0136] 290 levels

[0137] 301 First Method Steps

[0138] 302 Second Method Steps

Claims

1. A method for providing an artificial neural network (200), wherein, Perform the following steps: - Provide optical signals (100) for the network (200) so that the output content (210) of the network (200) can be obtained by processing the optical signals (100) by the network (200). - Using the spectral phase and / or temporal phase-specific properties of the optical signal (100) to provide at least one network component (250) of the network (200) for the processing, wherein the at least one network component (250) includes at least one neuron (251) and / or weight (252) of the network (200). - Provide multiple neurons (251) as the at least one network component (250), Its features are, The spectral combiner (284) of the network (200) is used to coherently or incoherently combine the output signals of the neurons (251) obtained by using the aforementioned characteristics in terms of spectral phase, in order to perform processing of multiple weights and / or neurons. The spectral combiner (284) is configured as a grating and / or prism sequence and / or dielectric layer and / or optical nonlinear medium and / or polarization optical device, and as the output signal of the spectral combiner (284), the spectral phase distribution map exists as a function of the frequencies of two individual neurons (251) and can be passed to the next neuron (251) through weights.

2. The method according to claim 1, Its features are, The optical signal (100) is provided by transmitting the optical signal to the network component (250), wherein the optical signal is a carrier of information and the information is processed by the network (200) to obtain an evaluation of the information as output content (210).

3. The method according to claim 2, Its features are, The information is processed by the network (200) to obtain a classification of the information as output content (210).

4. The method according to any one of claims 1 to 3, Its features are, The use of phase-specific properties of the optical signal (100) is achieved by non-linearly and linearly altering the spectral phase distribution of the optical signal (100) to change the properties in order to implement a non-linear function of the neuron (251) and a linear weighting of the weights (252).

5. The method according to any one of claims 1 to 3, Its features are, The output content (210) and / or the output signal (102) of at least one network component (250) are analyzed by means of analyzing the phase-specific characteristics of the optical signal (100).

6. The method according to any one of claims 1 to 3, Its features are, The optical signal (100) is provided in the form of an optical input signal (101), which is specific to the input information and is processed by the at least one network component (250) to obtain an optical output signal (102) specific to the output content (210).

7. The method according to any one of claims 1 to 3, Its features are, Multiple network components (250) are provided, the network components including neurons (251) and / or weights (252), the neurons and / or weights being arranged in different layers (290) of the network (200) and optically interconnected.

8. The method according to any one of claims 1 to 3, Its features are, The phase-specific properties of the optical signal (100) are either the spectral phase and / or the time phase itself, or the group delay, or the group delay dispersion, or the group velocity dispersion, or the third-order dispersion of the derivative of the spectral phase, or a higher-order dispersion.

9. The method according to any one of claims 1 to 3, Its features are, A network (200) is used in a vehicle (1), wherein the optical signal (100) is provided in the form of an optical input signal (101), which is specific to input information about the environment (2) of the vehicle (1) and is processed by the at least one network component (250) to obtain output content (210) as a classification of the input information.

10. A system for providing an artificial neural network (200), comprising: - An interface (270) for providing optical signals (100) for the network (200) so that the output content (210) of the network (200) can be obtained by processing the optical signals (100) by the network (200). - At least one optical element (230) for using characteristics of the optical signal (100) specific to the spectral phase and / or temporal phase of the optical signal (100) to provide at least one network component (250) of the network (200) for the processing, wherein, The at least one network component (250) includes at least one neuron (251) and / or weight (252) of the network (200). The network (200) includes multiple optical elements (230) designed to, for the purpose of utilizing the aforementioned characteristics, at least non-linearly alter the spectral phase distribution of the optical signal (100) to provide output signals (102) of neurons (251) of the network (200). The network (200) includes at least one spectral combiner (284) to spectrally combine the output signals (102) of the neurons (251), and a spectral phase analyzer (285) to analyze the phase of the combined output signals (102) to provide output content (210). The spectral combiner (284) is configured as a grating and / or prism sequence and / or dielectric layer and / or optical nonlinear medium and / or polarizing optical device, and is configured to perform the method according to any one of claims 1 to 9.

Images