DEMODULATING MODULATED SIGNALS WITH ARTIFICIAL NEURAL NETWORKS

Abstract

Demodulating a modulated signal. A method may include receiving a modulated signal, where the modulated signal is a signal modulated according to a modulation function that varies faster than the signal. The modulation function is a function of the signal. The received modulated signal is modulated by an artificial neural network (ANN) system trained to recognize bit values ​​from signal patterns generated by the modulation function by identifying bit values ​​from patterns of the received modulated signal. Related modulation and demodulation systems are disclosed.

Description

BACKGROUND

[0001] The present disclosure relates generally to demodulation techniques using artificial neural networks and in particular to techniques based on photonic data processing systems configured as optical reservoir networks.

[0002] Machine learning is mostly based on artificial neural networks (ANNs), which are computational models modeled on biological neural networks in the brains of humans or animals. These systems learn tasks step by step and independently using examples; they have been successfully used, for example, in speech recognition, text processing, and computer vision.

[0003] A KNN (Kinetic Neighborhood Network) has a set of connected units, or nodes, comparable to biological neurons in animal brains and therefore referred to as artificial neurons. Signals are transmitted along connections (also called edges) between artificial neurons, much like synapses. That is, an artificial neuron receiving a signal processes it and then sends signals to connected neurons. In implementations, the signals propagated along these connections are analogous real numbers, and the outputs of the artificial neurons are computed using a nonlinear function of the sum of their inputs. In photonic networks, signals can also be propagated as complex numbers.

[0004] Connection weights (also called synaptic weights) are typically associated with connections and nodes; these weights adapt during the learning process. Each neuron can have multiple inputs, and each input is assigned a connection weight (the weight of that specific connection). The training algorithm learns these connection weights during a training phase and is updated accordingly. The learning process is iterative: the network is typically presented with data cases one at a time, and the weights associated with the input values ​​are adjusted at each time step.

[0005] Many types of neural networks are known, ranging from feedforward neural networks like multilayer perceptrons, deep neural networks, and convolutional neural networks. New types of neural networks are also emerging, such as spiking neural networks. A spiking neural network (SNN) differs significantly from conventional neural networks because SNNs operate with spikes, which can be, for example, discrete binary events that occur asynchronously at any given time, rather than analog values ​​calculated at regular time intervals. This means that SNNs incorporate the concept of time in addition to the neuronal and synaptic state. Specifically, neurons fire only when a membrane potential reaches a certain value, instead of firing on every transmission cycle as in multilayer perceptron networks.In the context of SNNs, firing means that a neuron generates a signal that reaches other neurons, which in turn increase or decrease their potentials according to the signals they receive from other neurons.

[0006] Reservoir data processing systems are closely related and enable the analysis of dynamic input data by training the output for specific tasks, such as classification and prediction. The inputs are typically fed into a fixed, random dynamic system (the reservoir), and the reservoir's dynamics map the inputs to a higher-dimensional system. A suitably trained readout mechanism is then used to read the reservoir's state and map this state to the output, requiring training only on the readout stage, while the reservoir itself remains fixed. The main types of reservoir data processing include echo-state networks and liquid-state machines (which can be considered a special type of SNN).

[0007] Neural networks are typically implemented in software. However, a neural network can also be implemented in hardware, for example, as a resistive processing unit or as an optical neuromorphic system. Optical data processing systems (also called photonic data processing systems) are known, for instance, that rely on photons (generated, for example, by light sources such as lasers or diodes) for computation. Application-specific units such as optical correlators have been proposed, for example, that rely on optical data processing to detect and / or track objects or to classify serial optical time-domain data.

[0008] US Patent 2014 / 0376669A1 describes the use of a neural network in a receiver to distinguish a large number of input waveforms without using a very large number of conventional fitted filters. The neural network is trained under actual line conditions, unlike the ideal signal requirements of fitted filters. The finite waveforms are based on digital modulation principles. The best possible match is made between a received waveform from the noisy channel and the previously trained waveforms to extract data. A fitted filter based on a neural network enables the separate discrimination of data for each subcarrier channel in the receiver.

[0009] US 2018 / 0 232 574 A1 describes systems and methods for sending and receiving data using machine learning classification. In one example, a communications control system includes a processor, memory, and receiver. The processor receives a transmit signal using the receiver, extracts features from the transmit signal, classifies the transmit signal based on the extracted features, and, using a machine learning classifier, recognizes a message within the transmit signal based on the classified transmit signal. SUMMARY

[0010] According to a first aspect, the present invention is designed as a method for demodulating a modulated signal. In another aspect, the method includes receiving a modulated signal, wherein the modulated signal is a signal that is modulated according to a modulation function that varies faster than the signal (i.e., in the time domain). The modulation function is a function of the signal. That is, the signal can be considered a variable (i.e., argument or input) of the modulation function. The received modulated signal is modulated by an artificial neural network (ANN) system trained to recognize bit values ​​from signal patterns generated by the modulation function by identifying bit values ​​from patterns of the received modulated signal.

[0011] The present approach allows the received signal (e.g., received directly on the bitstream) to be decoded at a transmission rate higher than the bit rate of the bit sequence captured by the signal. Since an attacker does not know the rate used, it is practically impossible to reconstruct the bit sequence without a suitably trained system. Furthermore, in addition to the attacker not knowing the encoding rate and the initial bit rate, the resolution of the encoded signal may occur within the temporal resolution of the physical detection unit used by the attacker, making detection impossible for that unit.

[0012] In embodiments, the method further includes, prior to receiving the modulated signal: modulating the signal according to the modulation function to obtain the modulated signal; and transmitting the obtained modulated signal so that it can be received by the KNN system and subsequently demodulated.

[0013] In some embodiments, the signal to be modulated is a digital signal; this signal is modulated over each period corresponding to each of the discrete values ​​captured by the digital signal to obtain the modulated signal. Accordingly, the modulated signal can have variations (or oscillations) within the time interval corresponding to each discrete value of the initial signal, such that the resulting modulation pattern varies faster than the underlying bit sequence (the raw bit rate), and this occurs for each period corresponding to each discrete value of the initial digital signal.

[0014] In some embodiments, the modulation for each of the discrete values ​​involves modulating the digital signal according to the modulation function based on two or more discrete values ​​of the digital signal, wherein the discrete values ​​include each of the discrete values ​​as well as one or more previous discrete values ​​of the digital signal.

[0015] In some embodiments, the modulation of the digital signal involves transmitting the digital signal in a data stream to a modulator so that the modulator modulates the signal transmitted in the data stream during processing. The transmission then involves transmitting the modulated signal in a data stream to the KNN system so that the system demodulates the signal transmitted in a data stream, which it receives during processing. In some variants, demodulation occurs in a stream mode, while modulation is buffered.

[0016] In embodiments, the method further includes converting the received modulated signal into a discrete signal so that the KNN system demodulates the converted signal by recognizing bit values ​​from patterns of values ​​in the converted signal.

[0017] In some embodiments, the modulated signal is transmitted optically.

[0018] In some embodiments, the received modulated signal is an optical signal, and the KNN system is part of a photonic data processing system configured as a reservoir data processing system. In this case, the method may further include coupling the received modulated signal into the KNN so that the latter demodulates the coupled-in signal by recognizing bit values ​​from patterns of the coupled-in signal.

[0019] In some embodiments, the signal is modulated with an electro-optic modulator before being transmitted optically, so that the transmitted signal can be received by the photonic data processing system and subsequently demodulated.

[0020] In embodiments, the modulation of the signal involves modulating the amplitude and / or phase of a signal-carrying electromagnetic field.

[0021] In some embodiments, the modulation involves modulating two or more input signals according to the modulation function to obtain one or more modulated signals, each varying faster than the input signals. The modulation function is a function of the input signals, resulting in the subsequent reception of one or more modulated signals. That is, the modulated signals are received (at the receiver) after modulation. However, these signals could be modulated in parallel and thus received in parallel, for example, using multiple parallel transmission channels. In other variations, the modulated signals can be transmitted sequentially. In all cases, the one or more received signals are demodulated (e.g., during or after reception) using the KNN system by recognizing bit values ​​from patterns of the one or more received modulated signals.

[0022] In some embodiments, the KNN system is implemented as a trainable hardware unit, and the method further includes mapping temporal information captured from the received modulated signal onto one or more input nodes of an input layer of the KNN system prior to demodulation. The KNN system also has an output layer consisting of one or more output nodes. The input nodes of the input layer are connected to the output nodes of the output layer via connections. At least some of these connections are associated with adjustable weighting elements. To demodulate the modulated signal and thereby recognize bit values, the KNN system reads signals from the output nodes.

[0023] The modulated signal can be received, for example, as an optical input signal. In these cases, the KNN system can be implemented as part of a photonic data processing system configured as a reservoir data processing system, where the mapping of the temporal information involves coupling the optical input signal into an input node of the reservoir data processing system. The input node is connected via optical reservoir nodes of a reservoir layer of the reservoir data processing system to one or more optical output nodes. Each of the optical output nodes is connected via corresponding connections to one or more of the optical reservoir nodes. Each connection has an adjustable weighting element associated with it.

[0024] In embodiments, the signal before modulation encodes an n-ary code, where n is greater than or equal to two, and the modulated signal after modulation encodes an m-ary code, where m is generally greater than n.

[0025] In some embodiments, the method further includes training the KNN system so that it can recognize bit values ​​from signal patterns generated by the modulation function.

[0026] In some embodiments, after demodulating the modulated signal, the method further includes adjusting the modulation function based on the feedback obtained from the demodulated signal in order to adapt a property of a next modulated signal, as well as modulating a subsequent signal based on the adapted modulation function.

[0027] According to another aspect, the invention is designed as a demodulator for demodulating a signal. The demodulator particularly includes an input unit configured to receive a modulated signal, which is a signal modulated according to a modulation function that varies faster than the signal itself, the modulation function being a function of the signal. The demodulator further includes a k-nearest neighbors (KNN) system connected to the input unit so that the input unit couples the received signal into the KNN system during operation.In accordance with the present procedures, it is assumed that the KNN system is trained to recognize bit values ​​from signal patterns generated by the modulation function; the KNN system is configured to demodulate the coupled modulated signal by recognizing bit values ​​from patterns of the modulated signal it receives during operation.

[0028] In embodiments, the KNN system is a photonic data processing system configured as a reservoir data processing system, which is suitable for demodulating a modulated optical signal by detecting bit values ​​from patterns of the modulated optical signal during operation, which are received by the input unit and coupled into the KNN system.

[0029] In some embodiments, the KNN system is implemented as a trainable hardware unit within the demodulator. The KNN system comprises an input layer of one or more input nodes and an output layer of one or more output nodes, the input nodes being connected to the output nodes via links, and at least some of these links being associated with adjustable weighting elements. The KNN system is otherwise configured to recognize bit values ​​by reading signals from the output nodes.

[0030] In some embodiments, the KNN system is implemented as a photonic data processing system configured as a reservoir data processing system. This system has a single input node connected via optical reservoir nodes in a reservoir layer to one or more optical output nodes in the output layer. Each optical output node is connected to one or more of the optical reservoir nodes via corresponding connections. Each connection has an adjustable weighting element associated with it. Finally, the input unit is configured to receive the modulated signal as an optical input signal and couple it into the individual input node.

[0031] According to a further and final aspect, the present invention is designed as a modulation system for modulating and demodulating a signal. This system initially comprises a modulator configured to modulate a signal according to a modulation function in order to obtain a modulated signal. Here, too, the modulation function is a function of the signal, which varies more rapidly than the initial signal. The system further comprises a transmission unit functionally connected to the modulator to transmit a modulated signal received from it, and a demodulator as described above, wherein the input unit is configured to receive a modulated signal transmitted by the transmission unit during operation.

[0032] In embodiments, the signal to be modulated is assumed to be a digital signal, and the modulator is configured to modulate the digital signal over each period corresponding to each of the discrete values ​​captured by the digital signal in order to obtain the modulated signal.

[0033] In some embodiments, the modulator is further configured to modulate the digital signal for each of the discrete values ​​according to the modulation function based on two or more discrete values ​​of the digital signal, wherein the discrete values ​​include each of the discrete values ​​as well as one or more previous discrete values ​​of the digital signal.

[0034] In some embodiments, the modulator is suitable for modulating a signal transmitted in a data stream during processing, and the KNN system is further configured to demodulate a modulated signal that it receives during processing.

[0035] In some embodiments, the modulator and the KNN system form a photonic data processing system.

[0036] The following describes devices, computer-aided systems and methods that carry out the present invention, using non-limiting examples and with reference to the accompanying drawings. Brief description of different views of the drawings

[0037] The accompanying figures, in which identical reference numerals refer to identical or functionally similar elements in the individual views and which, together with the detailed description below, are included in the present description and form part thereof, serve to further illustrate various embodiments and to explain various basic ideas and advantages according to the present disclosure, in which: Fig. 1 is a diagram illustrating general steps of a method for modulating and demodulating signals according to embodiments; the Fig. Figures 2A to 2D and 3A to 3D are representations of examples of signals (or parts thereof) that occur in embodiments. These signals are represented in the time domain; the examples chosen are deliberately kept simple. Fig. Figure 2A shows an example of a discrete signal representing a sequence of values ​​{1, 0, 1, 0, 0, 1}, which is used as an input for modulation. Fig. Figure 2B illustrates an example of a synthesis signal that results from modulating the input signal of Fig. 2A is obtained with a modulation function, where the modulation function generates signal pulses that are in the Fig. 3A and Fig. 3C are represented based on a current (instantaneous) value (0 or 1) of the discrete input signal. Fig. 3B and Fig. 3D signals are discrete counterparts to the analog signal pulses of the... Fig. 3A and Fig. 3C. That in Fig. The signal received by 2B is indeed an analog signal, but the Fig. 2C and Fig. 2D shows examples of discrete modulated signals created by modulating the initial sequence of Fig. 2A can be obtained on the basis of advanced substitution schemes, where the modulation function uses previous values ​​of the signal as arguments in addition to an instantaneous value; Fig. Figure 4 schematically shows an optical reservoir network and selected components of a photonic data processing system configured to implement such a network as occurs in embodiments; Fig. 5 schematically illustrated selected components of another photonic data processing system, which is also configured as an optical reservoir as found in other embodiments; and Fig. 6 is a flowchart illustrating the general steps of a procedure for modulating and demodulating a signal according to embodiments.

[0038] The accompanying drawings show simplified representations of devices and systems, or parts thereof, that appear in various embodiments. The in Fig. The technical features shown in Figure 5 are not to scale. Unless otherwise stated, identical or functionally similar elements in the figures have been given the same reference numerals. DETAILED DESCRIPTION OF FORMATIONS OF THE INVENTION

[0039] Neural cryptography relies on stochastic algorithms (e.g., KNN algorithms) used in encryption and cryptanalysis. Using neural networks to process encrypted data is a viable option, as KNNs can, in principle, replicate any function. This means that a suitably trained KNN could potentially be used to find the inverse function of a cryptographic algorithm. While no practical applications have been proposed yet, it is known that encrypted data is vulnerable due to the rapid development of KNNs.

[0040] Based on this finding, the inventors have developed a conceptually simple solution for improving the security of encrypted data, which can advantageously utilize hardware-based encryption / decryption technology. This is explained in more detail in the following description, which is structured as follows. First, general embodiments and general variants are described (Section 1). The next section deals with more specific embodiments and details of the technical implementation (Section 2). 1. General designs and general variants

[0041] With regard to the Fig. 1 and Fig. Section 6 first describes an aspect of the invention which relates to a method for demodulating a modulated signal, i.e. a method which is implemented on the receiver side 20.

[0042] In some embodiments, this method is based on a modulated signal 54 received in step S12, where the modulated signal 54 is assumed to have been obtained by modulating (11, step S11) a given signal 51. This signal 54 was, in particular, modulated using a modulation function 52 that varies faster than the initial signal 51 (in the time domain). The modulation function 52 is a function of the signal 51: it uses the initial signal 51, its values, or values ​​captured therein as arguments to generate a modulated signal 53. Such a signal 53 can then be transmitted (as a signal 54) S12 and coupled as a signal 55 into an artificial neural network system 22 (hereinafter referred to as ANN system 22) for demodulation purposes S21.

[0043] This means that the modulated signal 55 is demodulated S22 using the KNN system 22, e.g., a trainable hardware unit or a computer system. More precisely, the system 22 is trained to recognize bit values ​​from signal patterns generated by the modulation function 52 during operation. The demodulation S22 is performed by recognizing bit values ​​from patterns of the received modulated signal 54.

[0044] The comments are listed in chronological order. In this context, the modulation / demodulation operations performed aim to encode (or encrypt) / decode (uncrypt) data or signal values ​​faster than the original signal. Demodulation can be performed, for example, directly on the signal itself (amplitude, phase, etc.), as explained later, or on the discrete values ​​it represents.

[0045] Signal 51, in its initial form (before modulation S11), can be, for example, a digital signal (e.g., as output by a digital circuit) or an analog signal (e.g., as output by a sensor or an analog circuit). A digital signal can be, for example, a pulse train (a pulse-amplitude modulated signal) or a physical signal that is sampled and quantized. A digital signal can therefore transmit a discrete signal (or discrete-time signal) as a time series of values. In all cases, the time-varying quantity in a digital signal is a representation of a sequence of discrete values ​​(a finite number of values). A digital signal is therefore often referred to as a discrete signal. In contrast, an analog signal is continuous-time, e.g.,A sinusoidal signal with a specific amplitude or phase modulation, where the time-varying quantity is a representation of another time-varying quantity. Regardless of whether it is an analog or digital signal, the initial signal 51 captures the information to be transmitted. This information can be captured, for example, as two or more bit values.

[0046] It is assumed that this initial signal 51 is modulated according to a modulation function f(.) S11. This function can be a digital modulation function that modulates discrete signals (as in the Fig. 3B, Fig. 3D illustrated) or continuous signals as output (as in the Fig. 3A, Fig. 3C). In all cases, the signal 53 resulting from modulation step S11 can be a digital or analog signal and can be forwarded (i.e., transmitted) S12, for example, electrically or optically. However, any wave transmission technique, such as acoustic waves, can be used in the intermediate transmission step S12. The changes in the resulting carrier signal caused by modulation S11 can, for example, form a finite (but usually large) number of alternative symbols, representing a modulation alphabet, as described in the embodiments below.

[0047] The modulated signal 54 received in step S12 can therefore be received in the form of a digital or analog signal. For example, it can initially be received as an analog signal S12, which is then, if necessary, converted into a digital signal S21 for subsequent demodulation purposes S22, e.g., sampled and interpreted as a set of consecutive discrete values. In this case, the KNN system S22 demodulates the converted signal by recognizing bit values ​​from value patterns in the converted signal S22. That is, recognition S22 is based on discrete values ​​and not on the properties of the signal. However, the demodulation step S22 does not necessarily have to be performed on discrete values, but can also be performed on the modulated signal itself, as mentioned above.This means that the recognition step S22 can be performed based on the properties of the signal, for example, based on an optical signal. The modulated signal can, for example, be transmitted optically and then optically coupled into the KNN system 22, which is configured as an optical reservoir as described in the embodiments below.

[0048] It should be noted that optical fiber data transmissions typically use square waves, and the underlying signal is usually considered a digital signal. However, in this case, the transmitted signal depends on the initial signal 51 and the modulation function 52 used, so the optical fiber signal may no longer be a square wave.

[0049] When a software implementation of KNN 22 is used, the demodulation step S22 directly yields a set of digits, e.g., {1, 0, 1, 0, 0, 1}, corresponding to the digits captured by the initial signal 51. In variants based on hardware implementations of KNN 22 (i.e., where the KNN is a dedicated hardware unit rather than a computer configured to implement a KNN), the demodulation step S21 typically yields an intermediate signal (not shown), from which the final sequence is derived as in Fig. 1 is assumed to be retrieved. For example, if an optical hardware network 22 is used, the signal received at the output of the KNN 22 can typically be an optical power stream (light with varying intensity) whose variations capture the same values ​​as the initial signal 51. Thus, the signal received as the output of step S22 can again be a discrete or analog signal, but one that captures values ​​corresponding to those of the initial signal 51. It should be noted that in embodiments, instead of a binary signal, notwithstanding the above, Fig. 1. Multi-stage signals can also be used in the representations used.

[0050] The demodulation step S22 can be viewed as a classification of patterns formed by the received signal S12 54, and subsequently, the signal S5 is coupled into the KNN 22 S21, with the classification being performed by the KNN system S22. The KNN system S22 is assumed to be trained to recognize bit values ​​from patterns generated by the modulation function S22, i.e., patterns of values ​​or corresponding signal patterns generated by this function S22 during operation. Thus, at inference time, the system S22 enables the recognition of bit values ​​from patterns of the modulated signal S5, either directly (as in software implementations of the KNN) or not (as in implementations using dedicated KNN hardware).

[0051] If necessary, the signal 54 received in step S12 can first be converted into digital values ​​S21 so that the system 22 can appropriately classify patterns formed in these digital values, i.e., values ​​that are already represented or captured by the signal 54. However, in one aspect, the KNN system performs operations directly on the received signal 54 via S22. In these cases, the signal 55 coupled into the KNN 22 is "identical" to the signal 54 received in step 54 (subject to the coupling step S21, which can affect the characteristics of the signal 54), so no analog-to-digital conversion is required in this case, making demodulation S22 more efficient. It should be noted that, for example, coupling the signal into a hardware KNN can degrade the signal, e.g., by causing optical losses and thus resulting in a smaller amplitude.Therefore, the coupled signal in this case is usually not completely identical to signal 54.

[0052] The present approach makes it possible to decode the received signal (e.g., received directly on the data stream) at a transmission rate higher than the bit rate of the bit sequence captured by signal 51. This is due to the fact that, as explained below, outputs from the modulation function vary more rapidly than the initial signal. Since an attacker does not know the modulation rate used, it is practically impossible to reconstruct the bit sequence without a suitably trained system. Furthermore, in addition to the attacker not knowing the encoding rate and the initial bit rate, the temporal resolution of the encoded signal 54 may fall within the resolution of the physical detection unit used by the attacker, making detection impossible for that unit.The use of a hardware-implemented KNN (Kinky-Nearest Neighbors) makes it possible, for example, to work with much higher frequencies than a conventional physical recognition unit. Such a KNN can, for instance, convert a "fast" encrypted signal into a "slow" decrypted signal.

[0053] The fact that the modulation function varies faster than the initial signal essentially means that, on average, the derivative of the modulated signal changes sign more frequently over the same period than the derivative of the initial signal (or a quantity captured from it). These conclusions hold true for continuous and differentiable signals and can also apply to higher (nth) order derivatives, for signals of the class Cn. For discrete / sampled signals, the fact that the modulation function varies faster means that, on average, the successive differences in the modulated signal values ​​change sign more frequently over the same period than the successive differences in the initial signal (or a quantity captured from it). These differences are calculated over intervals corresponding to the smallest available temporal resolution.Similar conclusions apply to the nth differences (n > 1). That is, the outputs of the modulation function oscillate faster in the time domain. This generally results in the modulated signal having a temporal frequency (i.e., one or more characteristic frequencies) that is, on average, higher than the temporal frequency (or characteristic frequencies) of the initial signal. Thus, the corresponding Fourier spectrum shifts to higher frequencies after modulation, and the autocorrelation function decays more rapidly from its maximum at lag 0 over the time interval corresponding to a minimum modulation period of the modulation function.

[0054] For discrete signals, the modulation function has a lower temporal resolution than the original signal. This is in Fig. The signal shown in 2C, for example, shows eight pulses, while the signal from Fig. 2A exhibits three pulses over the same time interval. The higher temporal frequency of the modulated signal can be characterized by conventional signal analysis and processing techniques, e.g., Fourier analysis, use of power spectra of the signals, etc. It should be noted that the above considerations refer to the useful parts of the original signal and the modulated signal after any noise has been filtered out. Regardless of any potential noise, the fact remains that the modulated signal varies more rapidly in its essential (and useful) part than the original signal in its essential (and useful) part.

[0055] As explained above, the KNN system 22 is a cognitive system that can be implemented in software or in hardware (as dedicated KNN hardware). The cognitive system can be implemented, for example, as any suitable machine learning model; that is, it can be implemented in software running on a conventional computer platform. Specifically, it can be implemented as a pulsed neural network or an autoencoder. Such a computer platform can nevertheless integrate dedicated accelerators and other hardware optimizations for KNNs. Likewise, co-optimized software and hardware platforms can also be used. The KNN system can also be implemented in dedicated KNN hardware, e.g.,as a trainable hardware unit, such as a photonic data processing system configured as an optical reservoir or a resistive processing unit, to increase demodulation speed and security for crypto applications. In particular, it can be advantageously implemented as an optical reservoir hardware system to enable high-speed decoding and crypto applications compatible with high raw bit rates based on optical signals, as described in detail later.

[0056] It should be noted that the terms cognitive algorithm, cognitive model, machine learning model, and similar terms are used interchangeably in the literature. To clarify the terminology, the following definition can be used provisionally: a machine learning model is generated by a cognitive algorithm that learns its parameters from input data points to arrive at a trained model. This allows us to distinguish between the cognitive algorithm being trained and the model that, after the underlying algorithm has completed its training, is ultimately referred to as the trained model. Similarly, a distinction can be made between a trainable, specialized KNN hardware and a trained hardware unit.It is assumed that the special hardware unit 22, which in some embodiments is used for demodulation purposes S22, has already been trained accordingly with respect to the modulation function 52, which was initially used to modulate S11 of the signal 51.

[0057] This will now be described in detail below with reference to certain embodiments of the invention.

[0058] The method and variants described so far relate to the decoding / demodulation step S22, which is performed at the receiver 20. However, the present invention comprises methods that are implemented both on the transmitter side 10 and on the receiver side 20. In embodiments, the present methods therefore also include modulating S11 of the initial signal 51 according to the modulation function 52 to obtain a modulated signal 53. The obtained modulated signal 53 is then transmitted S12 to be received by the KNN system 22 and subsequently demodulated S22.

[0059] The initial modulation S11 can be achieved, for example, by digital signal processing (DSP) if digital modulation is desired. The resulting S11-modulated signal 53 can then be transmitted electrically or optically S12 as explained above. In some variants, an electro-optic modulator (EOM) can be used to modulate a light beam S11, which can then be transmitted optically. A digital signal 51 can, for example, first be digitally processed and then optically encoded. For this purpose, fast EOMs such as lithium niobate EOMs or silicon-based EOMs, which are known per se, can be used. In other variants, the original (digital or analog) signal 51 can be subjected to analog signal processing S11.

[0060] Assume, for example, that the initial signal 51 is a digital signal as in Fig. 1 (see also) Fig. 2A). Since the modulation function 52 must vary faster than the input signal for the present purposes, the initial signal 51 can, for example, be modulated S11 over any period corresponding to each of the discrete values ​​captured by the digital signal 51 in order to obtain the modulated signal 53. Fig. Figures 2B to 2D show possible examples of resulting signals 53. Modulation S11, for example, can be based on instantaneous values ​​of the discrete signal 51, where the initial bit sequence 51 is transmitted in a data stream through modulator 11. As with amplitude modulation, an instantaneous value refers to the value currently being considered by function 52 of modulator 11 and in Fig. 1 is denoted by f(.). That is, the modulation function 52 can, based on a current value of the discrete signal and for each current value processed successively, generate a signal pulse (as in Fig. 2B) or generate a digital signal representing a set of several distinct values ​​(as in the Fig. 2C, Fig. 2D). As in the Fig. As shown in diagrams 2B to 2D, the modulated signal varies "faster" than the input signal. That is, the modulated signal always exhibits a change within the time interval corresponding to each discrete value of the initial signal. Fig. 2A corresponds to several variations (oscillations). All these time intervals are generally constant, so the resulting modulation pattern varies faster than the underlying bit sequence (the raw bit rate), and this is true for each period corresponding to each discrete value of the initial signal 51. It should be noted that the period over which the initial signal in Fig. 1A varies 1 (w. E.) is, while the basic period over which the signals of the Fig. 2C and Fig. 2D modulation is 1 / 4 (w. E.).

[0061] The Fig. 3A and Fig. Figure 3C shows two different parts of continuous (analog) signals generated by a given modulation function for converting a "1" and a "0". Applying this function to the initial sequence of Fig. When 2A is applied, this sequence is modulated and produces a signal like in Fig. 2B showed that the patterns according to the Fig. 3A and Fig. 3C entangled. Since the Fig. 3A and Fig. 3C analog-like conversions can be included, which is in Fig. The received signal from 2B is considered an analog signal.

[0062] In contrast, if a discrete modulation function were applied to the initial sequence of Fig. Applying 2A would result in a discrete signal. Fig. 3B and Fig. 3D, for example, are discrete counterparts to the analog parts of the Fig. 3A and Fig. 3C, which can be used to process the input signal from Fig. To modulate 2A. The result of such modulation is not shown.

[0063] It is understood, however, that the initial signal 51 may be a discrete signal modulated by a discrete or analog modulation function, or an analog signal sampled by a discrete modulation function, for example, for modulation purposes. In variations, analog modulation may be used to modulate the analog initial signal, for example, based on the instantaneous value of the read analog input signal (as in frequency modulation). Crucially for the present purpose, the modulation function must vary faster than the output signal (or the quantity represented by it).

[0064] The Fig. 2C and Fig. 2D instead shows modulated signals, which are created by modulating the initial sequence of Fig. 2A can be obtained according to advanced substitution schemes, in which previous values ​​of the initial signal are taken into account in addition to the instantaneous value. Fig. 2C is achieved in particular by modulating the initial sequence in Fig. 2A is obtained according to a substitution {k, l} → {p, q, r, s}, while Fig. 2D is obtained by modulating the same initial sequence according to the substitution {k, l} → {p, q, r}, which leads to different time step subdivisions. While the in Fig. Since the modulation performed by 2C remains binary, the modulation in Fig. 2D to a ternary code (multi-level signal). These substitution rules are described in detail below.

[0065] First, it should be noted that the modulation step S11 can be advantageously performed with operands consisting of several discrete values ​​of the initial signal 51. That is, each discrete value of the signal 51 can be modulated according to a modulation function 52 that takes two or more discrete values ​​of the signal 51 as inputs. For example, the function f can take, for each current value of the signal 51, the current value as well as one or more previous discrete values ​​of the signal as inputs, in order to, as in the Fig. 2C and Fig. 2D assumes generating a modulation. That is, the initial bit sequence {1, 0, 1, 0, 0, 1} is modulated according to a parametric modulation.

[0066] In the example of Fig. 2C is assumed to perform substitutions as listed below: {k, I} again denotes an input pair of values, where l is a current (instantaneous) value and k is the value immediately preceding the current value I in the initial bit sequence, while the sets {p, q, r, s} denote modulated sets that are generated at the output: {0,0}→{0,0,1,1} {0,1}→{0,1,1,0} {1,0}→{1,0,0,1} {1,1}→{1,1,1,0}

[0067] This means that the basic period is divided by two in this example (see below). Fig. 2C and Fig. 2A). It should be noted that these substitutions generally require proper initialization of the initial sequence to allow substitution for the first and last bits in the initial sequence. For example, a dummy bit (e.g., 0) can be prepended to the actual initial sequence. The in Fig. The initial sequence {1, 0, 1, 0, 0, 1} shown in Figure 2A can, for example, first be interpreted as the modified sequence {{0}, 1, 0,1, 0, 0, 1}, where {0} is prepended to {1, 0, 1, 0, 0, 1}. The first actual value of the sequence {1, 0, 1, 0, 0, 1} of Fig. 2A, which is 1, then yields an input pair value {0, 1} in accordance with the modified sequence. This in turn results in the modulation {0, 1, 1, 0}, which is performed in a time step twice as short as the time step of the initial sequence. The second actual value of the initial sequence {1, 0} of Fig. 2A is 0, resulting in the input pair {1, 0}, since the immediately preceding value is 1. This in turn leads to a modulated set {1, 0, 0, 1} according to the preceding substitution list, etc. This finally yields the result shown in Fig. The modulated pattern shown in Figure 2C is shown. The dummy bit prepended to the initial sequence must ultimately be removed from the discrete signal 56 once it has been reconstructed (after step S22). In some variants, a dummy bit can be appended instead of being prepended to the initial sequence.

[0068] It should be noted that when using discrete signals, neither the initial signal 51 nor the modulated signal 53 necessarily have to be binary signals (representing the values ​​of 0 or 1). The modulation function 52, for example, can use a ternary code, as in Fig. 2D assumed to encode information. However, the arity of the modulated code can exceed that of the input code to make it more difficult for an attacker to decode a sequence. Thus, while the output signal 51 can encode an n-ary code (with n ≥ 2), the modulated signal 53 after modulation S11 can encode an m-ary code with m > n. This is exemplified in Fig. 2D representation, where after modulation S11 a ternary code is obtained, in contrast to the binary encoding used at the input ( Fig. 2A).

[0069] More precisely, it is assumed that the following substitutions are performed to modify the signal from Fig. 2D after encoding S11. Here too, it is assumed that a dummy bit (0) has been prepended to the actual initial sequence. In the following list, {k, l} denotes an input pair of values, where l is the current value of the input signal, while k is a previous value of the given bit sequence, and {p, q, r} denotes modulated sets generated at the output of modulation function 52: {0,0}→{0,2,1} {0,1}→{0,1,2} {1,0}→{1,2,0} {1,1}→{2,1,0}

[0070] These substitutions lead to the in Fig. 2D represented ternary sequence.

[0071] Regarding variants Fig. In 2D (where the modulated signal encodes an m-ary code after modulation with m > n), the n-ary code can simply be converted into another n-ary code (i.e., m = n) based on a different temporal basis, as previously described with reference to Fig. Section 2C explains this. In other variants, the arity of the modulated signal can even be smaller than n (i.e., m < n), provided that the frequency of the signal is further modulated to compensate for the missing dimension. While the latter is easier to access, it can still be more stable in transmission.

[0072] It should be noted that the items in the Fig. The modulation examples shown in Figures 2B to 2D are for illustrative purposes only. In practice, the modulations performed in step S22 are usually carried out with much longer initial sequences {k, I, ...}, and longer output sequences are generated. Furthermore, the substitutions need not be static and can evolve over time. In addition, advanced dynamic coding may occur. Therefore, the primary KNN system 22 may need to be modified as shown in the flowchart of Fig. 6 assumed to be retrained S30 over time.

[0073] In simpler versions, however, each modulation step is based on, as in Fig. 2A is assumed to be based on a single value of the digital signal, i.e., its instantaneous value, where different bit values ​​correspond to different signal pulses (see the Fig. 3A, Fig. 3C) or discrete sequences ( Fig. 3B, Fig. 3D). Accordingly, the S22 modulation can potentially produce a continuously varying signal (as in Fig. 2B) or a discrete signal ( Fig. 2C, Fig. 2D). In all cases, the fact remains that the additional complexity of the higher frequency (or higher bit rate) of the output signal (or sequence) makes it very difficult to interpret the signal 54, 55 without a suitably trained KNN and prior knowledge of the time step (bit rate) of the initial signal 51 and the modulation function 52.

[0074] In embodiments, the digital signal 51 is initially transmitted in a data stream to a modulator 11 S10, so that the modulator 11 modulates the signal during processing. Here too, previous signal values ​​can be used together with the current signal value to perform the modulation S11 as described above by way of example. The modulated signal 54 can therefore be transmitted in a data stream to the KNN system 22, so that the system demodulates the signal it receives in a data stream during processing S22.

[0075] In variants, the values ​​of the signal could potentially be rearranged into blocks or block arrangements of bit sequences before the rearranged signals are modulated (S11). These sequences can then be modulated sequentially or in parallel (S11) and then demodulated in the same way before the value sequence is reassembled. In this respect, the modulation patterns can be learned over several parallel sets of modulated sequences (S30). Again, complex modulation schemes can be considered to form complex symbols, notwithstanding the simple examples in the Fig. 2 and Fig. 3.

[0076] One aspect involves modulating two or more input signals in parallel (using a modulation function 52, which is performed on all input signals considered) S11 to obtain one or more modulated signals 53. Again, the resulting modulated signal(s) vary faster than the input signals. The one or more modulated signals 54, received sequentially in step S12, are subsequently demodulated by the KNN system 22 through pattern recognition S22, i.e., by recognizing bit values ​​from patterns of the one or more modulated signals 54 received in step S12. The initial signals 51 can, for example, be input streams (i.e., sequences of data elements) in a software implementation of the KNN. In other variations, they can be a set of analog signals.The modulation step S11 can thus lead to a single current or a single modulated analog signal, or to multiple currents or signals, which are fed to the KNN for demodulation S21 in order to restore the information contained in the input signal(s).

[0077] The KNN can advantageously be a reservoir network. For reasons of efficiency and security, the KNN system 22 can now be implemented as a trainable hardware unit (e.g., dedicated KNN hardware) rather than in software. As explained above, the modulated signal 54 can also be transmitted primarily optically S12, where the signal 54 received in step S12 is an optical signal 54. In this case, a photonic data processing system 20 can advantageously be used to perform the demodulation S22. More precisely, the KNN system 22 can be part of a photonic data processing system 20 configured as a reservoir data processing system. Coupling S21 of the received optical signal 54 into the KNN 22 causes the KNN 22 to demodulate the coupled signal 55 S22 by recognizing bit values ​​from signal patterns in the coupled signal 55.

[0078] As previously described, signal 51 can, for example, be modulated by an EOM 11 S11 before being optically transmitted S12, so that the transmitted signal 54 is received by the photonic data processing system 20 S12 and subsequently demodulated S22. While an EOM can advantageously be used to modulate the signal, other solutions are also conceivable to achieve this. One such solution is to use a linear optical network with delay lines operating with coherent light; see, for example, https: / / arxiv.org / ftp / arxiv / papers / 1501 / 1501.03024.pdf.

[0079] For example, a linear optical network (with interconnected delay lines) can be used to encode information in both the amplitude and phase of the signal-carrying electromagnetic field. In some variants, the modulation S11 can be based solely on the light intensity (only the amplitude of the field is modulated) or solely on the phase. While the encoded signal 51 can be a digital signal, in other variants it can also be a purely analog signal (e.g., electrical), as described above.

[0080] Regardless of whether the KNN is implemented as a trainable hardware unit (i.e., dedicated KNN hardware) or in software, the fundamental working principle of the KNN remains the same, as now described with reference to... Fig. Section 4 described the same thing. First, the temporal information captured by the modulated signal 54 must be mapped to one or more input nodes 251 of the input layer of the KNN (S21). The input nodes 251 of the input layer are connected to the output nodes 254 of the output layer of the KNN via connections. At least some of these connections are associated with adjustable weighting elements, i.e., they can be adjusted during a training process, which makes the system 22 a trainable system. For the bit value recognition performed in step S22, signals must be read from the output nodes.

[0081] It should be noted that in Fig. 4. Several input and output nodes are assumed. In variants, however, each of the input and output layers can consist of a single node. In all cases, the temporal information (and possibly additional information) contained in the signal 54 received in step S12 can first be appropriately mapped to the input nodes 251 ...

[0082] It should be noted that a similar architecture can be implemented in software, for example, by a pulsed neural network or SNN, where output nodes are connected to each other via lateral all-to-all inhibitory connections, while input nodes are connected to output nodes via all-to-all excitatory connections, which have associated connection weights. Other types of KNNs, such as FFNs, can also be considered.

[0083] As described above, the modulated signal 54 received in step S12 can be an optical signal, and the KNN system 22 can advantageously be implemented as part of a photonic data processing system 20 configured as a reservoir data processing system. In this case, the temporal information can be mapped to the reservoir data processing system 22 S21 by coupling the optical signal 54 into its input node 251. As further described in Fig. As shown in Figure 4, an input node 251 can be connected via optical reservoir nodes 252 of a reservoir layer of the system 22 to one or more optical output nodes 254. Each of the optical output nodes 254 is typically connected via corresponding connections to one or more of the optical reservoir nodes 252, each of which has adjustable weighting elements 253. It should be noted that, notwithstanding the representation of the specific implementation of the system 22, the following applies: Fig. The 4 reservoirs shown are normally connected to node 252 in practice.

[0084] Reservoir data processing systems enable effective analysis of dynamic input data by training the output. The KNN system can be implemented, in particular, as a fluid state machine or an echo state network. In embodiments, the reservoir layer can be in the form of an optical interference pattern with a given optical power distribution, as in Fig. 5. This can be implemented. In this way, the temporal information of the optical input signal can be mapped onto the optical interference pattern. The output connections and associated weighting elements can also operate in the optical domain. The in Fig. The optical reservoir system shown in Figure 5 is further described in Section 2.1.

[0085] As in Fig. As shown in Figure 6, the present methods can further include training (S30) of the KNN system so that it recognizes bit values ​​from signal patterns generated by the modulation function 52. This can be achieved, in particular, by sending a training data set and subsequently training the KNN 22 based on the received data set (S30). The KNN 22 can be retrained (S30) as needed, for example, if the modulation conditions change or the signal path changes. In safety-sensitive applications, the training of the system 22 must be handled with care. For example, the training data sets can be transmitted using a different channel than the one used for inference (S22).It should be noted that when using an optical network 22, changing the physical connection path may require retraining of the system 22, especially if the phase of the signal is modulated S11 and / or interpreted S22.

[0086] The modulation function 52 used in step S11 may need to be adapted for safety reasons or to adapt to a dynamically evolving context S50: different data types used at a given time may require a different type of modulation. Furthermore, the modulation function 52 may need to be updated S50 based on the feedback S40 received from the KNN system 22 after the properties of the demodulated signal S12 have been analyzed S40, in order to adjust a property of a subsequently modulated signal 51 (a signal modulated during a subsequent cycle). This may be done in particular to meet or optimize certain boundary conditions (e.g., to maintain a constant mean value of the modulated signal), or with a view to improving signal transmission (e.g.,Reducing transmission errors by selecting an optimal function 52 to improve the physical contrast between different input states after decoding or the coding efficiency. It should be noted that the property(ies) analyzed in step S40 refer to a signal 55 received at the receiver. The analysis performed in step S40 thus affects the next cycle S11 to S22. The feedback obtained from step S40 can be used, for example, to ensure that a constant mean value of the modulated signal is maintained during subsequent cycles. It should be noted that updates to the modulation function 52 can be made immediately after each cycle S11 to S22 or after several cycles and several batches of training data.

[0087] A second trainable KNN (not shown) can be used, for example, for signal generation. This second KNN is connected upstream of the modulator 11 to influence the modulation function 52. Together with a control circuit on the receiver side 20, this second KNN can be used to adjust the coding function 52, for example, to meet or optimize boundary conditions.

[0088] According to other aspects, the invention can furthermore be implemented as a data processing system such as a demodulator 20 or an entire modulation system 1. Aspects of these systems have already been discussed implicitly in relation to the present methods and are only briefly mentioned below with reference to the Fig. 1, Fig. 4 and Fig. 5 described.

[0089] The invention can initially be implemented as a single demodulator 20 for demodulating a signal according to the methods described above. Such a demodulator 20 can comprise an input unit 21 and a KNN system 22. The input unit 21 is configured to receive a modulated signal 54, i.e., a signal that is modulated according to a modulation function 52 such that it varies more rapidly than the initial signal 51, as explained above. The KNN system is connected to the input unit 21 so that it couples the received signal 54 into the KNN system 22 during operation. As explained above, the KNN system 22 is assumed to be trained to recognize bit values ​​from signal patterns generated by the modulation function 52.The system 22 is otherwise configured in the demodulator 20 to demodulate modulated signals 55 coupled into it by recognizing bit values ​​from patterns of the received modulated signal 54.

[0090] With regard to the Fig. 4 and Fig. 5. The KNN system 22 can, in particular, be a reservoir data processing system, e.g., part of a photonic data processing system 20 configured as a reservoir data processing system. Such a reservoir data processing system is suitable for demodulating an optical signal 53, 54 by recognizing bit values ​​from patterns of the modulated optical signal 53, which is received by the input unit 21 and subsequently coupled into the KNN system 22 during operation.

[0091] As in Fig. As shown in Figure 4, the KNN system can, in particular, have an input layer of one or more input nodes 251 and an output layer of one or more output nodes 254. The input nodes 251 are connected to the output nodes via connections (arrows), with at least some of these connections belonging to adjustable weighting elements 253. The KNN system 22 can recognize bit values ​​by reading signals from the output nodes 254 during operation. More generally, however, the KNN system 22 can be implemented as a trainable, dedicated hardware unit in the demodulator 20.

[0092] The input unit 21 can further be configured to map temporal information (and / or specific information, other dimensional information such as wavelength, polarization, core in a multi-core optical waveguide) captured in the modulated signal 54 to input nodes of the system 22. In some variants, the input unit 21 can simply couple the signal 54 into a single input node (e.g., optically).

[0093] In the Fig. In the system 22 shown in Figure 4, the input nodes 251 are connected to the optical output nodes 254 via optical reservoir nodes 252 of a reservoir layer of the system 22. Each optical output node 254 is connected via corresponding connections to one or more of the optical reservoir nodes 252, each of which has adjustable weighting elements 253. The input unit 21 is configured to receive a modulated signal 54 as an optical input signal 54 and couples this ISi into the input nodes 251 as described in detail in Section 2.1.

[0094] Another example of a photonic data processing system 22a is in Fig. 5 is shown, which is described in detail in section 2.1.

[0095] It should be noted that in implementations based on optical reservoir networks, the receiver 20 may still include an optical detector, which, however, would be positioned downstream of the KNN 22. In other words, the signal 54 is first fed to the optical KNN 22 S21 and then detected by the optical detector, with the input unit 21 in this case essentially being a coupler that couples the received signal into the KNN.

[0096] In variants where a non-optical system such as an RPU is used instead of an optical reservoir, the input unit 21 should be configured (e.g. programmed) in a similar way so that it maps temporal information captured in the input signal (and possibly other information such as amplitude, phase, etc.) to different input nodes of the KNN so that these can forward the signals to higher layers of the network.

[0097] With renewed reference to Fig. 1. The invention can be implemented according to another aspect as a modulation / demodulation system 1 for modulating and demodulating a signal. That is, such a system 1 comprises a demodulator 20 (as described above) and a modulator 11. The latter is configured to modulate a signal 51 according to a modulation function 52 to obtain a modulated signal. As explained above, the modulation function 52 is a function of the signal 51 (i.e., it uses the signal 51 as input) and its outputs vary faster than the signal 51. The system 1 further comprises a transmission unit that is functionally connected to the modulator 11 to transmit modulated signals received from it.

[0098] Here too, the modulator 11 may be configured to modulate a digital signal 51, for example, over each period corresponding to each of the discrete values ​​captured by the digital signal 51, in order to obtain the modulated signal 53. Furthermore, the modulator 11 may be configured to modulate the digital signal 51 based on two or more discrete values ​​(e.g., the current value and two or more previous values) using a suitable modulation function 52. The modulator 11 may, for example, be suitable for modulating a signal transmitted in a data stream during processing. Likewise, the KNN system 22 may be designed to demodulate during processing, particularly when a special optical system is used.

[0099] The foregoing embodiments have been briefly described with reference to the accompanying drawings and may comprise a number of variants. Various combinations of the aforementioned features may be considered. Examples are described in the next section. 2. Specific Implementations - Technical Implementation Details 2.1 Optical Reservoir Data Processing Systems

[0100] Fig. Figure 4 schematically illustrates a photonic data processing system 22, which, according to embodiments, is implemented as an optical reservoir data processing system. The photonic data processing system 22 has an input layer, a reservoir layer, and an output layer. The input layer has a plurality of input nodes 251 configured to receive optical input signals ISi (e.g., optical input currents) and forward these signals ISi to the reservoir layer. The reservoir layer has a limited reservoir area, e.g., a continuous optical reservoir area 410 with a plurality of optical reservoir nodes 252. The reservoir layer is here represented as an optical interference pattern with an optical power distribution as shown below with reference to Fig. 5 described trained.

[0101] The photonic data processing system 22 further features a plurality of optical output connections between the reservoir nodes 252 and the output nodes 254. At least some of the optical output connections are associated with weighting elements (wi) 253, which can be adjusted during the training process S30. That is, the optical reservoir system 22 can be trained to perform specific computational tasks as described in Section 1.

[0102] In an optical reservoir system such as in Fig. As shown in Figure 4, the input nodes 251 can be designed as optical input waveguides, e.g. as coupling regions arranged at an intersection with the optical interference region (which forms the reservoir layer), while the reservoir nodes 252 can be designed as readout units.

[0103] The training process of the optical reservoir data processing system 22 changes the weights wi 253. However, the reservoir layer itself remains unchanged (other connections continue to have fixed weights that do not change during the training / learning process).

[0104] During operation, the output nodes 254 generate an optical output signal that can be converted into the electrical domain by suitable, known converters. The converted output signals can then be further processed in the electrical domain by suitable hardware or software processing means. The adjustment of the weights 253 can generally be performed in software or hardware. A hardware control circuit with additional control software running on it can, for example, receive the output signals of the output nodes 254 during the training process and adjust the weights of the optical weighting elements 253 by applying electrical control signals to the optical weighting elements 253. The optical weighting elements 253 can, for example, be implemented as optical attenuators or optical amplifiers. During the training process, certain states of the reservoir system can be evaluated.When using some learning algorithms, the state of the output connections 254 can be accessed according to the weighting elements 254. Parts of the optical signal can therefore be split and fed to a dedicated detector and the respective learning algorithm during the training process.

[0105] An optical reservoir data processing system 22 as in Fig. As shown in section 4, it can advantageously be operated according to the reservoir data processing paradigm.

[0106] In some embodiments, the opto-electrical conversion is performed at the optical output node 254. In other variants, however, this conversion can be performed upstream of node 254, for example, at the reservoir node 252. There, the output connections, the weighting elements 253, and the output node 254 can be implemented as electrical components. It should be noted that the optical reservoir itself and the reservoir node 252 remain exclusively within the optical domain. In other variants, the weighting elements 253 and the output node 254 can be implemented in software.

[0107] In all cases, the initial weights 253 can be trained to form a trained (or controlled) layer formed by the initial nodes 254, and subsequently derive results during an inference phase.

[0108] Fig. Figure 5 shows a schematic representation of another photonic data processing system according to embodiments.

[0109] The data processing system 22a has two feedback delay waveguides that are used to map temporal information contained in the optical input signal 54 onto the optical interference pattern 510. The optical interference pattern has an optical power distribution that represents the optical power at the respective locations of the optical interference pattern.

[0110] The data processing system 22a has a plurality of readout units 252, which, as with reference to Fig. 4 are described as reservoir nodes. The readout units 252 are located in an inner region of the optical interference area 510, in contrast to the edges 241 (grey shading in Fig. 5) of the optical interference region 510. This interior region can, for example, be defined as the entire area of ​​the optical interference region 510 with the exception of the outer edges 241. The edges 241 may be formed by a mirror structure, e.g., a Bragg reflector, or by a metal coating on the interference region 510. In some variants, the edges 241 may correspond to a transition between the interference region 510 and a surrounding region (not shown), which is formed, for example, by a layer (e.g., SiO2) with a different refractive index than the interference region 510 (e.g., Si).

[0111] The readout units 252 are configured to detect optical readout signals RSi of the optical power distribution at readout positions RPi in the interior of the optical interference area 510. The readout unit 252 can, for example, be configured to acquire optical intensities, optical powers, optical energies, and / or information about the optical phases.

[0112] The photonic data processing system 22a has two input delay waveguides 221 and 222. The input delay waveguide 221 serves to receive part of an input signal IS, e.g., via a splitter or a coupler. The input signal IS is then delayed by the input delay waveguide 221 and forwarded as the delayed input signal ISd1 with a first time delay d1 to the interference region 510. The delayed input signal ISd1 is then forwarded to the input delay waveguide 222, e.g., via a splitter or coupler.

[0113] The delayed input signal ISd1 is then further delayed by the input delay waveguide 222 and passed on to the interference region 510 as another delayed input signal ISd2 with a second time delay d2. Further combinations of feedback delay lines and input delay lines can be considered to achieve a desired temporal mapping.

[0114] System 22a can further incorporate nonlinear components, implemented, for example, as thermo-optic elements, electro-optic elements, electrical feedback loops, and / or optical resonators, to map the temporal information of the optical input signal onto the optical interference pattern. These nonlinear components provide a nonlinear dependence of the optical interference pattern on the optical input signals.

[0115] The optical data processing system 22a may, for example, have one or more nonlinear components 230, 242 to perform nonlinear signal conversion. The nonlinear components 242 may, for example, be arranged in the interference region 510 (as in Fig. 5 assumed). These nonlinear components 242 can be provided, for example, as a photorefractive element, optical amplifier or attenuator.

[0116] In some variants, nonlinear components can be arranged in an upstream input waveguide (not shown), the input delay waveguides 221, 222, and / or the feedback delay waveguides. For example, the optical resonators 230 can be located in the feedback delay waveguides 210 or the input delay waveguides 221, 222, as shown schematically in Fig.They may be arranged as shown in Figure 5. These resonators may be configured to have a finite optical lifetime.

[0117] In other variants, the interference region 510 in the optical data processing system 22a may possibly have one or more scattering elements (e.g. similar to the elements 242) to scatter the optical wave and increase the complexity of the optical interference pattern.

[0118] In other variants, the receiver may have 20 nonlinear elements that are not part of the optical system 22a itself. 2.2. Software Implementations of KNNs

[0119] As described in Sections 1 and 2.1, the KNNs can be implemented wholly or partially in software. The present invention can therefore be implemented as a system, a method, and / or a computer program product at any possible level of technical integration. The computer program product can comprise a computer-readable storage medium (or media) on which computer-readable program instructions are stored to instruct a processor to execute aspects of the present invention.

[0120] A computer-readable storage medium can be a physical unit capable of retaining and storing instructions for use by a unit to execute instructions. For example, a computer-readable storage medium can be an electronic storage unit, a magnetic storage unit, an optical storage unit, an electromagnetic storage unit, a semiconductor storage unit, or any suitable combination thereof, without limitation. A non-exhaustive list of more specific examples of computer-readable storage media includes the following: a portable computer disk, a hard disk, random-access memory (RAM), read-only memory (ROM), and erasable programmable read-only memory (EPROM).Flash memory), static random-access memory (SRAM), portable compact storage disk-read-only memory (CD-ROM), DVD (digital versatile disc), USB flash drive, floppy disk, a mechanically coded unit such as punched cards or raised structures in a groove on which instructions are stored, and any suitable combination thereof. A computer-readable storage medium shall not, in its use herein, be understood as volatile signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses traveling through an optical fiber), or electrical signals transmitted by a wire.

[0121] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to individual data processing units or, via a network such as the internet, a local area network, a wide area network, and / or a wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission lines, wireless transmission, routers, firewalls, switching units, gateway computers, and / or edge servers. A network adapter card or network interface in each data processing unit receives computer-readable program instructions from the network and forwards them for storage on a computer-readable storage medium within the respective data processing unit.

[0122] Computer-readable program instructions for executing work steps of the present invention may be assembler instructions, ISA (Instruction Set Architecture) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, configuration data for integrated circuits, or either source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., as well as procedural programming languages ​​such as the programming language "C" or similar programming languages.The computer-readable program instructions can be executed entirely on the user's computer, partially on the user's computer as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, via the internet using an internet service provider).In some embodiments, electronic circuits, including, for example, programmable logic circuits, field programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), can execute computer-readable program instructions by using state information from the computer-readable program instructions to personalize the electronic circuits to implement aspects of the present invention.

[0123] Aspects of the present invention are described herein with reference to flowcharts and / or block diagrams or diagrams of methods, devices (systems), and computer program products according to embodiments of the invention. It is pointed out that each block of the flowcharts and / or block diagrams or diagrams, as well as combinations of blocks in the flowcharts and / or block diagrams or diagrams, can be executed by means of computer-readable program instructions.

[0124] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a specialized computer, or another programmable data processing device to create a machine such that the instructions executed by the processor of the computer or other programmable data processing device generate a means of implementing the functions / steps specified in the block(s) of the flowcharts and / or block diagrams or charts.These computer-readable program instructions may also be stored on a computer-readable storage medium capable of controlling a computer, programmable data processing device, and / or other units to function in a particular manner, such that the computer-readable storage medium on which instructions are stored has a manufacturing item, including instructions that implement aspects of the function / step specified in the block(s) of the flowchart and / or block diagrams or charts.

[0125] The computer-readable program instructions can also be loaded onto a computer, other programmable data processing device, or other unit to cause the execution of a series of process steps on the computer or other programmable device or other unit in order to generate a process executed on a computer, such that the instructions executed on the computer, other programmable device, or other unit implement the functions / steps specified in the block(s) of the flowcharts and / or block diagrams or charts.

[0126] The flowcharts and block diagrams or charts in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this context, each block in the flowcharts or block diagrams or charts can represent a module, segment, or part of instructions that includes one or more executable instructions for performing the specific logical function(s). In some alternative embodiments, the functions specified in the block may occur in a different order than shown in the figures. For example, two blocks shown consecutively may in reality be executed essentially simultaneously, or the blocks may sometimes be executed in reverse order depending on the corresponding functionality.It should also be noted that each block of the block diagrams or charts and / or flowcharts, as well as combinations of blocks in the block diagrams or charts and / or flowcharts, can be implemented by special hardware-based systems that perform the specified functions or steps, or execute combinations of special hardware and computer instructions.

[0127] Although the present invention has been described with reference to a limited number of embodiments, variants, and the accompanying drawings, it is understood by those skilled in the art that various modifications can be made and equivalents substituted without departing from the scope of the present invention. In particular, a feature illustrated in a particular embodiment, variant, or drawing (which is equivalent to a unit or a method) can be combined with or replaced by another feature in a different embodiment, variant, or drawing without departing from the scope of the present invention. Therefore, various combinations of the features described with respect to one of the foregoing embodiments or variants can be considered that still fall within the scope of the accompanying claims.Furthermore, many minor modifications can be made to adapt a particular situation or material to the teachings of the present invention without deviating from its scope. It is therefore intended that the present invention is not limited to the specific embodiments disclosed, but rather encompasses all embodiments falling within the scope of the appended claims. Moreover, many other variants besides those explicitly described above are conceivable.

[0128] Typical combinations of features disclosed herein are shown in the following clauses: 1. A method for demodulating a modulated signal, wherein the method comprises: Receiving a modulated signal, wherein the modulated signal is a signal that is modulated according to a modulation function that varies faster than the signal, the modulation function being a function of the signal; and Demodulating the received modulated signal using an artificial neural network system (ANN system) trained to recognize bit values ​​from signal patterns generated by the modulation function by recognizing bit values ​​from patterns of the received modulated signal. 2. Method according to clause 1, wherein the method continues to have the following before receiving the modulated signal: Modulating the signal according to the modulation function to obtain the modulated signal; and Transmitting the received modulated signal so that it can be received by the KNN system and subsequently demodulated. 3. Procedure according to clause 2, wherein the signal is a digital signal; and The signal is modulated over each period of time corresponding to each of the discrete values ​​that are captured from the digital signal to obtain the modulated signal. 4. Method according to clause 3, wherein the modulation of the digital signal includes: for each of the discrete values, a modulation of the digital signal according to the modulation function based on two or more discrete values ​​of the digital signal, wherein the discrete values ​​include each of the discrete values ​​as well as one or more previous discrete values ​​of the digital signal. 5. Procedure according to clause 3, wherein The modulation of the digital signal involves transmitting the digital signal in a data stream to a modulator so that the modulator can modulate the signal transmitted in a data stream during processing, and The transmission involves transferring the modulated signal in a data stream to the KNN system so that it can demodulate the signal it receives in a data stream during processing. 6. Procedure according to clause 1, wherein The procedure further involves converting the received modulated signal into a discrete signal so that the KNN system can demodulate the converted signal by recognizing bit values ​​from patterns of values ​​in the converted signal. 7. Method according to clause 2, wherein the modulated signal is transmitted optically during transmission. 8. Procedure according to clause 7, wherein the received modulated signal is an optical signal and the KNN system is part of a photonic data processing system configured as a reservoir data processing system, and The method further involves coupling the received modulated signal into the KNN so that it demodulates the coupled signal by recognizing bit values ​​from patterns of the coupled signal. 9. Procedure according to clause 8, wherein the signal is modulated with an electro-optical modulator before it is optically is transmitted so that the transmitted signal can be received by the photonic data processing system and subsequently demodulated. 10. Procedure according to clause 9, wherein The modulation of the signal involves modulating the amplitude and / or phase of a signal-carrying electromagnetic field. 11. Procedure according to clause 2, wherein The modulation comprises the modulation of two or more input signals according to the modulation function in order to obtain one or more modulated signals, each varying faster than the input signals, wherein the modulation function is a function of the input signals such that one or more modulated signals are subsequently received; and The one or more received signals are demodulated using the KNN system by recognizing bit values ​​from patterns of the one or more received modulated signals. 12. Procedure according to clause 1, wherein the KNN system is implemented as a trainable hardware unit, The method, prior to demodulating the received modulated signal, further comprises mapping temporal information captured from the received modulated signal onto one or more input nodes of an input layer of the KNN system, wherein the KNN system further comprises an output layer of one or more output nodes, the input nodes of the input layer being connected to output nodes of the output layer via connections, at least some of these connections belonging to adjustable weighting elements, and The recognition of the bit values ​​involves reading signals from the output nodes. 13. Procedure according to clause 12, wherein the modulated signal is received as an optical input signal, and the KNN system is implemented as part of a photonic data processing system configured as a reservoir data processing system, wherein the mapping of the temporal information involves coupling the optical input signal into an input node of the reservoir data processing system, wherein the input node is connected via optical reservoir nodes of a reservoir layer of the reservoir data processing system to one or more optical output nodes of the output layer, Each of the optical output nodes is connected via appropriate connections to one or more of the optical reservoir nodes, and Each connection has its own adjustable weighting elements. 14. Procedure according to clause 1, wherein The signal before its modulation encodes an n-ary code, where n is greater than or equal to two, and the modulated signal after its modulation encodes an m-ary code, where m is generally greater than n. 15. Procedure according to clause 1, wherein The procedure further involves training the KNN system so that it can recognize bit values ​​from signal patterns generated by the modulation function. 16. Procedure according to clause 2, wherein the procedure further comprises: After demodulating the modulated signal, the modulation function is adjusted based on the feedback received from the demodulated signal to adapt a property of the next modulated signal, and Modulating a subsequent signal based on the adapted modulation function. 17. Demodulator for demodulating a signal, wherein the demodulator comprises: an input unit configured to receive a modulated signal, where the signal is modulated according to a modulation function that varies faster than the signal, the modulation function being a function of the signal; and an artificial neural network (ANN) system connected to the input unit so that it couples the received signal into the ANN system during operation, the ANN system being trained to recognize bit values ​​from signal patterns generated by the modulation function and configured to demodulate the coupled-in modulated signal by recognizing bit values ​​from patterns of the received modulated signal. 18. Demodulator according to clause 17, wherein The KNN system is a photonic data processing system configured as a reservoir data processing system, capable of demodulating a modulated optical signal by recognizing bit values ​​from patterns of the modulated optical signal received by the input unit and coupled into the KNN system during operation. 19. Demodulator according to clause 17, wherein the KNN system is implemented as a trainable hardware unit in the demodulator, wherein the KNN system has an input layer of one or more input nodes and an output layer of one or more output nodes, wherein the input nodes are connected to the output nodes via connections and at least some of these connections are associated with adjustable weighting elements, and wherein the KNN system is configured to recognize bit values ​​by reading signals from the output nodes. 20. Demodulator according to clause 19, wherein the KNN system is implemented as a photonic data processing system configured as a reservoir data processing system, having a single input node connected via optical reservoir nodes of a reservoir layer to one or more optical output nodes of the output layer, Each of the optical output nodes is connected via appropriate connections to one or more of the optical reservoir nodes, Each connection has its own adjustable weighting elements, and The input unit is configured to receive the modulated signal as an optical input signal and couple it into the individual input node. 21. Modulation system for modulating and demodulating a signal, wherein the system comprises: a modulator configured to modulate a signal according to a modulation function to obtain a modulated signal, where the modulation function is a function of the signal that varies faster than the signal, a transmission unit that is functionally connected to the modulator in order to transmit the modulated signal received from it, a demodulator that features: an input unit configured to receive a modulated signal during operation, transmitted by the transmission unit; and an artificial neural network (ANN) system connected to the input unit so that the input unit couples the received signal into the ANN system during operation, wherein the ANN system is trained to recognize bit values ​​from signal patterns generated by the modulation function and is configured to demodulate the coupled-in modulated signal during operation by recognizing bit values ​​from patterns of the received modulated signal. 22. Modulation system according to clause 21, wherein the signal is a digital signal and the modulator is configured to modulate the digital signal over each period of time corresponding to each of the discrete values ​​captured by the digital signal in order to obtain the modulated signal. 23. Modulation system according to clause 22, wherein the modulator is still configured to output the following for each of the discrete values The digital signal is modulated according to the modulation function based on two or more discrete values ​​of the digital signal, wherein the discrete values ​​include each of the discrete values ​​as well as one or more previous discrete values ​​of the digital signal. 24. Modulation system according to clause 21, wherein the modulator is suitable for modulating a signal transmitted in a data stream during processing, and The KNN system is configured to demodulate a modulated signal it receives during processing. 25. Modulation system according to clause 21, wherein the modulator and the KNN system constitute a photonic data processing system.

Claims

[1] Method for demodulating a modulated signal, wherein the method comprises: Receiving (S12) a modulated signal (53, 54, 55), wherein the modulated signal is a signal that is modulated according to a modulation function (52) that varies faster than the signal, the modulation function being a function of the signal; and Demodulation (S22) of the received modulated signal using an artificial neural network system, KNN system (22), wherein the latter is trained to recognize bit values ​​from signal patterns generated by the modulation function by recognizing bit values ​​from patterns of the received modulated signal, wherein the KNN system is implemented as a trainable hardware unit, The method further comprises, prior to demodulating the received modulated signal, a mapping (S21) of temporal information captured from the received modulated signal onto one or more input nodes (251) of an input layer of the KNN system, wherein the KNN system further comprises an output layer of one or more output nodes (254), the input nodes of the input layer being connected to output nodes of the output layer via connections, at least some of these connections being associated with adjustable weighting elements (253), and The recognition of the bit values ​​involves reading signals from the output nodes, whereby the modulated signal is received as an optical input signal, and the KNN system is implemented as part of a photonic data processing system configured as a reservoir data processing system, wherein the mapping of the temporal information involves coupling the optical input signal into an input node of the reservoir data processing system, wherein the input node is connected via optical reservoir nodes (252) of a reservoir layer of the reservoir data processing system to one or more optical output nodes of the output layer, Each of the optical output nodes is connected via appropriate connections to one or more of the optical reservoir nodes, and Each connection has its own adjustable weighting elements. [2] Method according to claim 1, wherein the method further comprises, prior to receiving (S12) the modulated signal (53, 54, 55): Modulating (S11) the signal according to the modulation function (52) to obtain the modulated signal; and Transmitting the received modulated signal so that it can be received by the KNN system and subsequently demodulated. [3] Method according to claim 2, wherein the signal is a digital signal (51); and the signal is modulated during modulation (S11) over each period corresponding to each of the discrete values ​​that are captured by the digital signal to obtain the modulated signal (53, 54, 55). [4] Method according to claim 3, wherein the modulation (S11) of the digital signal (51) comprises: for each of the discrete values ​​a modulation of the digital signal according to the modulation function (52) based on two or more discrete values ​​of the digital signal, wherein the discrete values ​​include each of the discrete values ​​as well as one or more previous discrete values ​​of the digital signal. [5] Method according to claim 3, wherein The modulation (S11) of the digital signal (51) comprises a transmission (S10) of the digital signal in a data stream to a modulator (11) so that the latter modulates the signal transmitted in a data stream during processing, and the transmission (S12) comprises a transmission of the modulated signal in a data stream to the KNN system (22) so that the latter demodulates (S22) the signal transmitted in a data stream that it receives during processing. [6] Method according to claim 1, wherein the method further comprises converting the received modulated signal (53, 54, 55) into a discrete signal (56) so that the KNN system (22) demodulates (S22) the converted signal by recognizing bit values ​​from patterns of values ​​in the converted signal. [7] Method according to claim 2, wherein the modulated signal (53, 54, 55) is transmitted optically during transmission (S12). [8] Method according to claim 7, wherein the received modulated signal (53, 54, 55) is an optical signal and the KNN system (22) is part of a photonic data processing system configured as a reservoir data processing system, and the procedure further involves coupling the received modulated signal into the KNN so that it demodulates the coupled signal (S22) by recognizing bit values ​​from patterns of the coupled signal. [9] Method according to claim 8, wherein the signal is modulated with an electro-optic modulator (11) before being optically transmitted so that the transmitted signal (53, 54, 55) is received by the photonic data processing system and subsequently demodulated (S22). [10] Method according to claim 9, wherein the modulation (S11) of the signal (51) comprises modulation of an amplitude and / or a phase of a signal-carrying electromagnetic field. [11] Method according to claim 2, wherein the modulation (S11) comprises modulating two or more input signals according to the modulation function (52) to obtain one or more modulated signals (53, 54, 55) which each vary faster than the input signals, wherein the modulation function (52) is a function of the input signals such that one or more modulated signals are subsequently received; and the one or more received signals are demodulated (S22) by the KNN system (22) by recognizing bit values ​​from patterns of the one or more received modulated signals. [12] Method according to claim 1, wherein the signal (51) before its modulation (S11) encodes an n-ary code, where n is greater than or equal to two, and the modulated signal (53, 54, 55) after its modulation encodes an m-ary code, where m is generally greater than n. [13] Method according to claim 1, wherein the method further comprises training (S30) of the KNN system (22) so that it recognizes bit values ​​from signal patterns generated by the modulation function (52). [14] The method of claim 2, wherein the method further comprises: After demodulating (S22) the modulated signal (53, 54, 55), an adjustment (S50) of the modulation function (52) is performed based on the feedback obtained from the demodulated signal (56) in order to adapt a property of a next modulated signal, and Modulating (S11) a subsequent signal based on the adapted modulation function. [15] Demodulator for demodulating (S22) a signal, wherein the demodulator has: an input unit (21) configured to receive a modulated signal (53, 54, 55) (S12), which is a signal modulated according to a modulation function (52) that varies faster than the signal, the modulation function being a function of the signal; and an artificial neural network system, KNN system (22), which is connected to the input unit so that the latter couples the received signal into the KNN system during operation, wherein the KNN system is trained to recognize bit values ​​from signal patterns generated by the modulation function and is configured to demodulate the coupled-in modulated signal (S22) by recognizing bit values ​​from patterns of the received modulated signal, wherein the KNN system is implemented as a trainable hardware unit in the demodulator, wherein the KNN system has an input layer of one or more input nodes (251) and an output layer of one or more output nodes (254), wherein the input nodes are connected to the output nodes via connections and at least some of these connections are associated with adjustable weighting elements (253), and wherein the KNN system is configured to recognize bit values ​​by reading signals from the output nodes, wherein the KNN system is implemented as a photonic data processing system configured as a reservoir data processing system having a single input node connected via optical reservoir nodes (252) of a reservoir layer to one or more optical output nodes of the output layer, Each of the optical output nodes is connected via appropriate connections to one or more of the optical reservoir nodes, The respective connections each have adjustable weighting elements, and The input unit is configured to receive the modulated signal as an optical input signal and couple it into the individual input node. [16] Demodulator according to claim 15, wherein the KNN system (22) is a photonic data processing system configured as a reservoir data processing system, wherein the latter is suitable for demodulating a modulated optical signal (53, 54, 55) by detecting bit values ​​from patterns of the modulated optical signal received by the input unit (21) and coupled into the KNN system during operation. [17] Modulation system for modulating and demodulating a signal, wherein the system comprises: a modulator (11) configured to modulate a signal (51) according to a modulation function (52) to obtain a modulated signal (53, 54, 55), wherein the modulation function is a function of the signal that varies faster than the signal, a transmission unit (12) which is functionally connected to the modulator, to transmit the modulated signal received from it, the demodulator according to claim 15 or 16, wherein the signal received by the demodulator is the modulated signal. [18] Modulation system according to claim 17, wherein the signal (51) is a digital signal and the modulator (11) is configured to modulate the digital signal over each period corresponding to each of the discrete values ​​that are detected by the digital signal in order to obtain the modulated signal (53, 54, 55). [19] Modulation system according to claim 18, wherein the modulator (11) is further configured to modulate the digital signal (51) for each of the discrete values ​​according to the modulation function (52) on the basis of two or more discrete values ​​of the digital signal, wherein the discrete values ​​comprise each of the discrete values ​​as well as one or more previous discrete values ​​of the digital signal. [20] Modulation system according to claim 17, wherein the modulator (11) is suitable to modulate (S11) a signal transmitted in a data stream (S10) during processing, and the KNN system (22) is configured to demodulate (S22) a modulated signal (53, 54, 55) that it receives during processing. [21] Modulation system according to claim 17, wherein the modulator (11) and the KNN system (22) are a photonic data processing system.

Images