Photonic circuit architecture for reservoir computing

Abstract

The present invention relates to a photonic circuit architecture for analogue reservoir computing, the architecture comprising: - an input for an input optical signal; - a reservoir circuit (24) capable of receiving a first intermediate signal (S̃) dependent on the input optical signal and of converting the first intermediate signal (S̃) into a converted signal, the reservoir circuit (24) being formed by coupled optical resonators (42), the reservoir circuit (24) being devoid of delay lines between the optical resonators (42); - a readout circuit capable of receiving a second intermediate signal dependent on the converted signal and of generating an output optical signal according to the second intermediate signal; and - an output for the output optical signal.

Description

[0001] DESCRIPTION

[0002] Photonic circuit architecture for reservoir computing

[0003] The present invention relates to a photonic circuit architecture for reservoir analog computing. The present invention also relates to a method for non-linear filtering of an optical signal implemented by such an architecture.

[0004] Externally emitted analog signals are generally not directly usable for calculations or analysis. Consequently, processing is required to eliminate noise, compensate for distortions, and extract relevant information.

[0005] These operations are usually performed in the digital domain, by linear and non-linear filters and, since the development of machine learning, by neural networks. For example, filtering and other linear operations are classically performed to decode advanced modulation formats in telecommunications (but there is also the field of radar signal analysis, free-space communications, sensors in general, and still other fields). Distortion is corrected using fast, but very power-consuming, digital electronics. In the conventional approach, neural networks are of the feed-forward type through a large number of layers. Thanks to the back propagation algorithm, they can learn a large number of complex tasks.These neural networks are based on traditional computing architectures (Von Neumann) and therefore consume a significant amount of energy, especially during the learning phase. In order to reduce energy consumption, an analog implementation of neural networks is considered.

[0006] In analog computing, the laws of physics are used to implement specific mathematical operations. Therefore, an analog computer has the advantage of potentially consuming less power than a digital computer. Furthermore, depending on the physical law considered, an analog computer can also be faster. In addition, the time lag between signal input and output, i.e., latency, is generally minimized.

[0007] Analog computing can be implemented in a variety of ways, particularly along the lines of a reservoir computer. A reservoir computer is a sufficiently complex nonlinear dynamic system whose internal state is described by the values ​​of certain physical quantities, such as the electric field, at fixed positions in space, called internal nodes. The evolution of a reservoir computer is governed by parameters that dictate the connection between nodes. The output is generated by connecting the internal nodes of the reservoir to output nodes.

[0008] The reservoir computer is robust with respect to manufacturing tolerances. Training is simplified in Echo State Networks (ESNs) because only external connection nodes are trained.

[0009] ESNs are governed by the following equation: = (1 - «)*ni + a tanh(l „[l; ï^] 7 ' + Wx„-i) (1 ) where:

[0010] • is the state of the network at time n,

[0011] • is the state of the network at time n-1,

[0012] • a is a parameter between 0 and 1.

[0013] • î n is a matrix mapping the inputs onto the state .

[0014] • is the input signal at time n,

[0015] • W is the reservoir weight matrix, transforming state n into state n+1 before nonlinear transformation (hyperbolic tangent y=tanh(x)).

[0016] The output signal of an ESN network is a linear combination of the states of the neurons, namely: yn ^out- -n

[0017] Or :

[0018] • w out is a matrix mapping the state to the outputs.

[0019] The simplicity of reading allows the use of pseudo-orthogonal regression (in English "ridge regression") to perform fast and easy learning that always converges. Another advantage is that ESN is robust to variation in system parameters.

[0020] Alongside the introduction of ESNs, liquid state machines (LSMs) were proposed. Conceptually, they are similar to ESNs, but they were designed from the ground up for hardware implementation, which also implies a mathematical description where time is a continuous variable, as opposed to ESNs whose state evolves in discrete time intervals. The two concepts were later grouped together and called reservoir computing (RC).

[0021] In the field of photonics, RC considers the propagation of light in a chip based on waveguides and other components, which is called an integrated photonic circuit. In the state of the art, implementations of reservoir computing are carried out by propagating light between "nodes". A variant consists of injecting the signal into a single node which is reconnected to itself by an optical delay line. In known implementations in photonics, part of the operations described by equation 1 is carried out by electronics.

[0022] However, in both cases, this involves the use of very fast electronics or long delay lines to ensure the separation of signals into time slots. Such delay lines are bulky and generate additional losses.

[0023] There is therefore a need for an implementation of reservoir computing with a compact architecture while limiting losses.

[0024] For this purpose, the present description relates to a photonic circuit architecture for analog reservoir computing, the architecture comprising: an input for an optical input signal, a reservoir circuit suitable for receiving a first intermediate signal dependent on the optical input signal and for converting the first intermediate signal into a converted signal, the reservoir circuit being formed of coupled optical resonators, the reservoir circuit being devoid of delay lines between the optical resonators,

[0025] - a reading circuit capable of receiving a second intermediate signal dependent on the converted signal and of generating an optical output signal as a function of the second intermediate signal and

[0026] - an output for the optical output signal.

[0027] Depending on specific implementation methods, the architecture includes one or more of the following characteristics, taken in isolation or in all technically possible combinations:

[0028] - each optical resonator is directly coupled to its near neighbors with a coupling rate and / or to a waveguide with a coupling rate;

[0029] - each optical resonator has an optical resonance frequency corresponding to the resonant mode, the optical resonators having a non-linear response so that the optical resonance frequency of each optical resonator depends on the energy stored in this same optical resonator;

[0030] - the signal propagating between the input and output of the architecture is only an optical signal, without conversion into the electrical domain;

[0031] - the reservoir circuit and the reading circuit are each produced in the form of an integrated photonic circuit;

[0032] - the output optical signal is a coherent sum of the optical fields in the optical resonators of the tank circuit;

[0033] - the reading circuit comprises additional outputs for additional signals, the additional signals being obtained by the reading circuit as a function of the second intermediate signal;

[0034] - the reading circuit comprises N or N-1 phase modulators and N-1 variable separators connected to the phase modulators;

[0035] - each variable separator is implemented by means of two phase modulators in a Mach-Zehnder interferometer;

[0036] - the architecture comprises: a first optical amplification and filtering circuit connected between the input and the reservoir circuit, the first optical circuit being capable of generating the first intermediate signal from the input optical signal, and a second optical amplification and filtering circuit connected between the reservoir circuit and the reading circuit, the second optical circuit being capable of generating the second intermediate signal from the converted signal.

[0037] The present invention also relates to a method for non-linear filtering of an optical signal implemented by an architecture as described previously.

[0038] Other characteristics and advantages of the invention will appear on reading the following description of embodiments of the invention, given by way of example only and with reference to the drawings which are:

[0039] - figure 1, a schematic representation of an example of a photonic circuit architecture for reservoir computing, the architecture comprising a reservoir circuit and a readout circuit,

[0040] - figure 2, a schematic representation of an example of implementation of the reservoir circuit of figure 1, - figure 3, a schematic representation of an example of the reading circuit of figure 1,

[0041] - figure 4, a schematic representation of an example implementation of the reading circuit of figure 3, the reading circuit comprising phase modulators and variable separators connected to the phase modulators,

[0042] - figure 5, a schematic representation of an example of a phase modulator of figure 4, and

[0043] - figure 6, a schematic representation of an example of a variable separator of figure 4.

[0044] A photonic circuit architecture 10 for reservoir computing is illustrated in Figure 1. Reservoir computing is mentioned in the introduction. More generally, reservoir computing is a computational framework derived from recurrent neural network theory that projects one or more input signals into higher-dimensional computational spaces by means of the dynamics of a fixed, nonlinear system, called a reservoir. Once the input signal is fed into the reservoir, which is treated as a "black box", a readout mechanism is trained to read the reservoir state and projects it to the desired output.

[0045] In particular, the architecture is capable of transforming an optical signal according to a law that the architecture (reservoir part) learns from training input and output sequences.

[0046] In particular, the architecture's reservoir (reservoir circuit + reading circuit) operates in two modes: "training" and "inference". In the first mode (training), the architecture 10 is presented with an input signal S and an expected output Yexpect, this is compared with the output Y of the reading circuit and the reservoir's Wout parameters are varied until the error is minimized. In the "inference" phase, the parameters W out are frozen.

[0047] Such an architecture 10 is, for example, suitable for use as a non-linear filter on optical signals, for example telecom signals, or radar signals on optical carrier. In a particular application, the architecture 10 is used for correcting distortions of telecom signals.

[0048] In the example illustrated by Figure 1, the architecture 10 comprises an input 20 for an input optical signal S, a first optical amplification and filtering circuit 22, a reservoir circuit 24, a second optical amplification and filtering circuit 26, a reading circuit 28 and an output 30 for an output signal O. Note that the first optical circuit 22 and the second optical circuit 26 are optional. In the example of Figure 1, an electronic block 32 is also illustrated, used, on the one hand, during a preliminary training phase aimed at setting the parameters of the reading circuit 28, and on the other hand, in order to control the first optical circuit 22, the reservoir circuit 24 and the second optical circuit 26. The electronic block 32 is optional during the inference phase, that is to say once the training phase is finished and the parameters of the reading circuit 28 are set.

[0049] Input 20 is suitable for receiving the input optical signal S. The input optical signal S has a bandwidth BW, a carrier frequency v0 and an average optical power P m .

[0050] The first optical amplification and filtering circuit 22 is connected between the input 20 and the reservoir circuit 24. The first optical circuit 22 is capable of amplifying and filtering the optical signal S to obtain a first intermediate signal S. The amplifying part can be produced using integrated photonic technology in indium phosphide and coupled to a silicon or silicon nitride photonic circuit performing the filtering.

[0051] In an exemplary implementation illustrated by FIG. 1, the first optical circuit 22 comprises an amplification module 34 with a gain G and a filtering module 36 (low-pass filter for example). The gain G makes it possible to bring the optical power back to the optimal level for the operation of the reservoir circuit 24. The bandwidth of the BWF filter is adjusted to eliminate the spectral components outside the signal band S. The first optical circuit 22 may consist of several amplifier-filter stages.

[0052] When the architecture 10 is devoid of a first optical circuit 22, the first intermediate signal S is the first optical signal S.

[0053] The tank circuit 24 is suitable for carrying out the calculation function per tank.

[0054] The reservoir circuit 24 is capable of receiving the first intermediate signal S and of converting the first intermediate signal S into a converted signal U. The reservoir circuit 24 forms the reservoir of the architecture 10.

[0055] In an exemplary implementation, the reservoir circuit 24 is capable of transforming the first intermediate signal S into the converted signal U on the basis of the following equation:

[0056] Or :

[0057] • a is a vector representing the internal state of the tank circuit. Each component a m is a complex number such that the resonant electric field at optical frequency in each cavity m contained in the reservoir is written E m (r, t) = a m (tyijj(f), and • f p is a function (more precisely a non-linear operator C M x C L -> C M, where M is the number of resonators and L the number of signals) modeling the evolution of the field U in the reservoir forced by the signal S.

[0058] The converted signal U is proportional to a subset of N resonators such that Ui = a^Ku, with K Ü the coefficients representing the coupling of resonator i with guide I.

[0059] This formalism corresponds to the approximation commonly used in the field of photonics and which is called "coupled-mode theory" [H. Haus, et al. "Coupled-mode theory." Procs. IEEE 79.10 (1991): 1505-1518.]

[0060] The reservoir circuit 24 is formed of coupled optical resonators 42. The optical resonators 42 are, for example, distributed in an arrangement of MIXM2 resonators (Figure 2). The optical resonators are also called optical cavities, or simply cavities, in the following.

[0061] The reservoir circuit 24 is devoid of delay lines between the optical resonators 42. An optical delay line is for example a waveguide whose length is such that the propagation time of the light is of the order of T = 100 ps. In the intended applications, the length is of the order of 1 cm. A waveguide of length of "1 mm is not a delay line, in the operational context considered. The different optical resonators 42 are not connected to each other by delay lines.

[0062] The 24 tank circuit is made entirely from optical components.

[0063] For example, the reservoir circuit 24 is produced using photonic technology on silicon, silicon nitride, lithium niobate, or even indium phosphide.

[0064] Resonators are, for example, silicon rings placed on a low-index substrate, for example silica. The rings are formed by looping waveguides. Waveguides are, for example, silicon ribbons whose width is around 500 nm (between 300 nm and 100 nm) and whose height is between 100 and 600 nm (depending on the materials). The geometry (height / width) is chosen according to the materials in order to minimize losses (attenuation by diffusion) and control the number of transverse modes, subject to the standards imposed by the foundries. The radius of the ring, or an alternative shape (e.g. "circuit") is chosen in order to minimize insertion losses and fix the free spectral range (ISL), in English "Free Spectral Range" (FSR). In our case, ISL is chosen "BW", which fixes the maximum value of the radius R = v g / (2TT*ISL), where v g is the group velocity of light in the waveguide, typically v g~ O.25*co, with Co the speed of light in vacuum. The distance between rings is comparable to the wavelength (~1.5pm), typically between 200 nm and 2 pm, depending on the technology, which induces optical coupling between two (or more) resonators. An example of a chain of coupled resonators is described here [F. Xia, et al., "Ultra-compact high order ring resonator filters using submicron silicon photonic wires for on-chip optical interconnects," Opt. Express 15, 1 1934-1 1941 (2007)]

[0065] Alternatively, resonators can be realized by approaching optical cavities of the "photonic crystal" type, as described in this publication [E. Yüce, et al. "Adaptive control of necklace states in a photonic crystal waveguide." ACS photonics 5.10 (2018): 3984-3988.].

[0066] In an exemplary implementation illustrated by Figure 2, a reservoir architecture based on ring resonators (or racetrack or micro-disk type) organized according to a matrix is ​​shown. Each optical resonator 42 (or optical cavity) is coupled to the near neighboring optical resonators directly according to a mutual coupling ratio p and some optical resonators 42 are coupled to a waveguide 44 with a coupling ratio K. The coupling ratio K expresses the decay rate of the energy stored in the optical resonator 42 due to the flow of a portion thereof in the waveguide. The coupling ratios K and p decrease when the distance between rings and rings and waveguide increases. This distance is therefore chosen in order to obtain the desired value of the coupling.In particular, in this example, one waveguide is used to connect to the first intermediate signal S and all others are associated with N outputs. Thus, the waveguides are only used to route the signal either to one optical resonator 42 or to one output, and not between the optical resonators 42.

[0067] There are, however, variations: for example, one could have several S signals at the input, routed by different waveguides. Note also that only a fraction of resonators (preferably those on the edge of the matrix) are coupled to waveguides.

[0068] Each optical resonator 42 (denoted by the index m) has an optical resonance frequency v m corresponding to the resonant mode whose amplitude is proportional to the variable a m. We only consider one mode of this optical resonator 42. Preferably, the other modes of the optical resonator 42 are spaced further apart than the bandwidth BW of the signal considered. Preferably, the frequency spacing of the modes in the same resonator (ISL) is larger than the bandwidth BW of the signal considered.

[0069] The optical resonators 42 are directly coupled to each other according to their spatial proximity with a mutual coupling ratio p (Figure 2, sub-Figure a). In particular, the spacing between the optical resonators 42 is less than the wavelength of the optical signal considered. Note that in these resonators the modes are twice degenerate, because at the same frequency correspond two modes according to the direction of propagation. The arrows show the direction of the modes which are coupled. Note that the architecture takes into account the direction of propagation of the light in each resonator, in particular with regard to the arrangement of the waveguides.

[0070] Subfigures 2a and 2b show the coupling between the optical resonators 42, carried out in an evanescent manner, i.e. the optical field extends over several optical resonators 42, thus forming "supermodes". The energy flow from one resonator to another can be considered as a "jump" with a delay T d = 1 / |_i, hence inversely proportional to the mutual coupling rate p between the optical resonators 42. This delay is also related to the time scale at which the optical field changes inside each resonator, i.e. to the optical bandwidth. This can be expressed by a delay-bandwidth product (a dimensionless figure related to the "memory" of the system). For example, in a single resonator, this product is equal to 1 . In a system of resonators, it can be greater than 1 .

[0071] Optical resonators 42 can be made with photonic crystals [D. Dodane, et al. "Fully embedded photonic crystal cavity with Q= 0.6 million fabricated within a full-process CMOS multiproject wafer." Optics express 26.16 (2018): 20868- 20877.]

[0072] This type of resonator is represented by squares (Figure 2b, c) and the modes are not degenerate. It is useful, in some cases, to couple these resonators by a very short waveguide (Figure 2c).

[0073] In one example, the optical resonators 42 have a non-linear response such that the resonant optical frequency v of each optical resonator 42 depends on the energy E stored in the optical resonator 42. The frequency v is, for example, given by the following equation (the index m is omitted for brevity): v = v0+ / 3. E (3)

[0074] Or :

[0075] • v0 is the frequency of the resonator at the limit of very weak optical excitation,

[0076] • p represents the strength of the nonlinear effect, and

[0077] • E is the energy in the resonator.

[0078] The frequency change is represented by a complex number, which means that p is also a complex constant. Physically, the effect described is either a frequency shift, a change in linewidth (increased absorption losses), or both. By weak excitation we mean that the spectral shift or spectral broadening is comparable to the width of the resonance Av = v0 / Ç of the optical resonator 42, where Q is the overpotential factor.

[0079] Preferably, all the optical resonators 42 are identical. They therefore have in this case the same resonant frequency (within the limits of manufacturing tolerances). Alternatively, at least one optical resonator 42 is different from the other optical resonators 42.

[0080] Preferably, the reservoir circuit 24 is produced in the form of an integrated photonic circuit. The integrated photonic circuit comprises, for example, at least one of the following materials: silicon, silicon nitride, silicon-rich oxide, III-V semiconductor-on-insulator alloy, and lithium niobate.

[0081] In a variant, it is possible to dynamically control the frequency v0 of each resonator 42 by thermo-optical, electro-optical or piezo-optical effect. The implementation depends on the chosen photonic foundry and in any case is a function proposed by it. Thus, some or all of the resonators 42 will have electrical terminals associated with them, such that the change in frequency Av dy " <x I ou V est proportionnel à un courant ou à une tension. Typiquement le terminal de masse est commun, ce qui réduit le nombre de terminaux à N terminaux < M + 1.

[0082] The second optical amplification and filtering circuit 26 is connected between the reservoir circuit 24 and the reading circuit 28. The second optical circuit 26 is suitable for amplifying and filtering the converted signal U to obtain a second intermediate signal Ü.

[0083] In an exemplary implementation illustrated by FIG. 1, the second optical circuit 26 comprises an amplification module 44 with a gain G and a filtering module 46 (low-pass filter for example).

[0084] When the architecture 10 is devoid of a second optical circuit 26, the second intermediate signal Ü is the converted signal U.

[0085] The reading circuit 28 is suitable for performing the reading function of the calculation by reservoir. The reading circuit 28 has, thus, been previously trained to read the state of the reservoir circuit 24 according to the principle of calculation by reservoir.

[0086] Preferably, the reading circuit 28 is produced in the form of an integrated photonic circuit. The integrated photonic circuit comprises, for example, at least one of the following materials: silicon, silicon nitride, silicon-rich oxide, III-V semiconductor-on-insulator alloy, and lithium niobate. The reading circuit 28 is capable of receiving the second intermediate signal Ü and of generating the output optical signal O as a function of the second intermediate signal Ü. In an exemplary implementation, the output optical signal O is obtained by a coherent sum of the optical fields in the optical resonators 42 of the reservoir circuit 24. In other words, the output optical signal O is obtained directly on the basis of the optical fields, without going through the intensities of these fields (which would have required the presence of a detector).More precisely, the reading circuit 28 is capable of performing a weighted sum of the components of the intermediate signal Ü (matrix) to generate the output optical signal O.

[0087] Mathematically, the reading circuit 28 realizes the function O = W 0Ut U, where U is the output (we omit the tilde sign for brevity) and O represents the field leaving the circuit (through waveguides) such that:

[0088] E W gf, t) = O (t)ll W g ( ) (4) and W out is an M to 2 matrix for example (so O is a 2-dimensional vector representing two outputs, or two modes in the same waveguide). This is different from an incoherent sum readout circuit, where the output signal is in the electrical domain and is written as O eiect (t) = m c m \ ^ m \ 2 et it is only the sum of positive components.

[0089] Preferably, as illustrated by the example of Figure 3, the reading circuit 28 comprises additional outputs for additional signals U x i, ... , U X N-2. Additional signals U x i, ... , U X N-2 are obtained by the read circuit 28 at the same time as the outputs O. For example, Figure 4 shows a photonic circuit diagram realizing the transformation described by equation 4. The additional outputs are obtained from the unused outputs of the components a.

[0090] Additional U signals x i, ... , U X N-2 allow more information to be extracted regarding the state U of the tank 24. The additional signals U x i, ... , U XN-2 are, for example, used to implement control actions of the architecture 10, for example actions aimed at adjusting parameters of the first optical circuit 22 and / or of the reservoir circuit 24 and / or of the second optical circuit 26.

[0091] An example of implementation of the reading circuit 28 is described in application WO 2017 / 144895.

[0092] In an exemplary implementation illustrated by FIG. 4, the reading circuit 28 comprises N (or N-1 ) phase modulators 50 and N-1 variable separators 52 connected to the N (or N-1 ) phase modulators 50. The phase modulators 50 and the variable separators are for example produced in photonic technology on silicon, thin-film lithium niobate on oxide [D. Zhu, et al., "Integrated photonics on thin-film lithium niobate," Adv. Opt. Photon. 13, 242-352 (2021 ); Qi, Yifan and Li, Yang. "Integrated lithium niobate photonics" Nanophotonics, vol. 9, no. 6, 2020, pp. 1287-1320]. The different components are made in the same photonic chip and connected by waveguides, represented by lines in Figure 4.

[0093] Figure 5 illustrates the definition of a phase modulator through the operation performed on the electric field E. Phase modulators in integrated optics exploit the electro-optical, piezoelectric, thermo-optical effect. In semiconductors (silicon, indium phosphide) the modulation of the density of free carriers is also exploited. Reference [Zhu2021, cited] describes the realization of phase modulators in lithium niobate. The realization of phase modulators in other platforms is described previously.

[0094] Figure 6 illustrates an example of a variable splitter 52. In this example, each variable splitter 52 is implemented by means of two phase modulators 60 in a Mach-Zehnder interferometer (MZI).

[0095] In particular, the implementation of the reading circuit 28, as presented in Figures 5 and 6, makes it possible to minimize the number of optical components used, i.e. 2N-1 , where N represents the number of outputs used. Moreover, one of the phase modulators 50 can be omitted, the total number of components becomes 2(N-1) . Other configurations are possible, for example by a chain of MZI in "push-pull" mode [Po Dong et al., "Highspeed low-voltage single-drive push-pull silicon Mach-Zehnder modulators," Opt. Express 20, 6163-6169 (2012)]. It is important to note that the modulation function is used in quasi-static conditions, i.e. at very low speed. Therefore, thermooptic effects can be suitable in the embodiment described here [K. Suzuki, et al, “Ultra-high-extinction-ratio 2 x 2 silicon optical switch with variable splitter,” Opt. Express 23, 9086-9092 (2015)].The preferred embodiment favors an embodiment aimed at minimizing the power consumption and the size of the device, for example as in the device described herein [Z. Han, et al. "High-performance and power-efficient 2x 2 optical switch on silicon-on-insulator." Optics Express 23.19 (2015): 24163-24170.].

[0096] The output 30 of the architecture 10 is the output of the reading circuit 28. In particular, in the inference phase, the output signal O is intended to be sent to the user, while during the learning phase, the output signal O and its complement Ô are sent to a detector (electronic block 32).

[0097] An example configuration of the architecture 10 will now be described. In this example, the electronic block 32 is used to drive the reading circuit 28, i.e. to determine the parameters of the reading circuit 28 for reading the output of the reservoir circuit 24. In this example, the electronic block 32 comprises a detector-receiver 50, a comparator 52 and a computer 54 (for example a digital electronic circuit). The detector-receiver transforms the optical intensity into fast electrical signals (bandwidth >1 GHz) and digitizes it. Advantageously, the detector-receiver 50 is a coherent optical receiver, which detects both the amplitude and the phase of an optical signal and digitizes them.

[0098] In this example, the output signal O (in a time interval T) is sent to the detector-receiver 50. The comparison between the expected signal Yexpect and the obtained signal is done in the digital domain (by the computer 54) and the result is the error £. In this case, the block 52 is included in the computer 54.

[0099] It is also possible to perform the error calculation in the analog electrical domain. In this case, the detector 50 converts the output signal O (optical signal) into the electrical domain and sends it, together with the expected signal Yexpect, to an electronic comparator-integrator circuit generating an error signal E.

[0100] The error signal E is then sent to the computer 54 to optimize the parameters of the first optical circuit 22 (gain G), the reservoir circuit 24, the second optical circuit 26 (gain G) and the reading circuit 28 (matrix M).

[0101] In one example, in the training phase, the procedure performs an iterative type error minimization algorithm and consists of the following:

[0102] 1 The error c(i) is measured for the response O(i) to the signal S(i), i indicating a realization of the signal in the interval Ti.

[0103] 2 the calculator 54 modifies the parameters W out in the reading circuit 28 as a function of the error signal £(i), and with the aim of minimizing the error signal E. For example, the new parameters are calculated according to the fastest descent algorithm.

[0104] 3 Steps 1 and 2 are repeated with new realizations of the S signals.

[0105] 4 Steps 1, 2, 3 are repeated but this time with the aim of optimizing the parameters of the tank circuit 24 and the value of the gain G.

[0106] 5 The auxiliary signals D are obtained by direct detection of the auxiliary outputs U X=VU by integrated low-frequency photodiodes (<1 GHz). These signals are used by the computer 54 to stabilize the operating point of the tank circuit 24. For example, it is possible to set the average power level to be maintained in each auxiliary output U as a setpoint. x . This function is also useful in the inference phase. An example of how architecture 10 works in the inference phase will now be described (inference phase).

[0107] Input 20 receives an optical signal from input S.

[0108] Optionally, the input optical signal S is amplified and filtered by the first optical circuit 22 to obtain the first intermediate signal S.

[0109] The first intermediate signal S is received by the reservoir circuit 24 and is converted into a converted signal U.

[0110] Optionally, the converted signal U is amplified and filtered by the second optical circuit 26 to obtain the second intermediate signal Ü.

[0111] The second intermediate signal Ü is received by the reading circuit 28 which generates the optical output signal O.

[0112] Thus, the architecture 10 is entirely implemented so that during the inference phase, no digital / electronic circuit is used to convert the optical signals. As a result, the signal propagating between the input 20 and the output 30 of the architecture 10 is only an optical signal, without conversion into the electrical or digital domain.

[0113] Therefore, in the inference phase, the energy consumption is essentially linked to 1) the amplification of the signals in the circuits 22 and 26, 2) the maintenance of the configuration in the reservoir circuit 24 and in the reading circuit 28; the computer 54. This consumption is relatively independent of the bandwidth of the optical signal, when in the digital algorithms this is a strongly increasing function of the bandwidth (of the flow rate).

[0114] If the phase modulators 50 and the power dividers 52 exploit the piezoelectric or electro-optical effect, their consumption can be extremely low.

[0115] Another advantage of all-optical operation in "inference" mode is that the circuit can be easily reconfigured to perform different tasks: correction of distortion of telecom signals coded according to different modulation formats and bit rates. The same architecture 10 can also be used to predict the future evolution of an optical signal generated by a complex system, for example a chaotic one.

[0116] The absence of delay lines in the tank circuit 24 allows a drastic reduction in the size of this part of the architecture 10. In the prior state of the art, a 16-node tank requires a 16 mm circuit 2[K. Vandoorne, et al. "Experimental demonstration of reservoir computing on a silicon photonics chip." Nature communications 5.1 (2014): 3541.], i.e. 3 orders of magnitude larger than the size of the reservoir 24 in the same photonic technology (silicon on oxide). This makes it possible to considerably increase the number of "neurons" in the reservoir. In order to exploit this advantage and translate it into computing power, a larger number of reservoir outputs must be used. The limit to the number of outputs U is given by the complexity of the readout circuit 28. It is possible to consider a readout circuit also made with resonators as for example in the reference [Dan Yi, et al., "Multi-functional photonic processors using coherent network of micro-ring resonators". APL Photonics 6, 100801 (2021).].

[0117] As a result, architecture 10 has a very small footprint (by several orders of magnitude if the reading circuit is also made with resonators) and reduced losses compared to state-of-the-art systems using delay lines.

[0118] Those skilled in the art will understand that the embodiments and variants of the description may be combined provided that they are technically compatible.

Claims

CLAIMS 1. Architecture (10) of photonic circuits for analog reservoir computing, the architecture (10) comprising: an input (20) for an input optical signal (S), a reservoir circuit (24) capable of receiving a first intermediate signal (S) dependent on the input optical signal (S) and of converting the first intermediate signal (S) into a converted signal (II), the reservoir circuit (24) being formed of coupled optical resonators (42), the reservoir circuit (24) being devoid of delay lines between the optical resonators (42), - a reading circuit (28) capable of receiving a second intermediate signal (Ü) dependent on the converted signal (U) and of generating an optical output signal (O) as a function of the second intermediate signal (£7), and - an output for the optical output signal (O).

2. Architecture (10) according to claim 1, wherein each optical resonator (42) is directly coupled to its near neighbors with a coupling rate (p) and / or to a waveguide (44) with a coupling rate (K).

3. Architecture (10) according to claim 2, wherein each optical resonator (42) has a resonant optical frequency (v) corresponding to the resonant mode, the optical resonators (42) having a non-linear response so that the resonant optical frequency (v) of each optical resonator (42) depends on the energy stored in this same optical resonator (42).

4. Architecture (10) according to any one of claims 1 to 3, in which the signal propagating between the input (20) and the output of the architecture (10) is only an optical signal, without conversion into the electrical domain.

5. Architecture (10) according to any one of claims 1 to 4, in which the reservoir circuit (24) and the reading circuit (28) are each produced in the form of an integrated photonic circuit.

6. Architecture (10) according to any one of claims 1 to 5, in which the output optical signal (O) is a coherent sum of the optical fields in the optical resonators (42) of the reservoir circuit (24).

7. Architecture (10) according to any one of claims 1 to 6, in which the reading circuit (28) comprises additional outputs for additional signals (U x i, U X N-2), the additional signals (U x i, U X N-2) being obtained by the reading circuit (28) as a function of the second intermediate signal (Ü).

8. Architecture (10) according to any one of claims 1 to 7, in which the reading circuit (28) comprises N or N-1 phase modulators (50) and N-1 variable separators (52) connected to the phase modulators (50).

9. Architecture (10) according to claim 8, wherein each variable separator (52) is implemented by means of two phase modulators (60) in a Mach-Zehnder interferometer.

10. Architecture (10) according to any one of claims 1 to 9, wherein the architecture (10) comprises: a first optical amplification and filtering circuit (22) connected between the input (20) and the reservoir circuit (24), the first optical circuit (22) being capable of generating the first intermediate signal (S) from the input optical signal (S), and a second optical amplification and filtering circuit (26) connected between the reservoir circuit (24) and the reading circuit (28), the second optical circuit (26) being capable of generating the second intermediate signal (Ü) from the converted signal (U).

11. Method for non-linear filtering of an optical signal implemented by an architecture (10) according to any one of claims 1 to 10.

Images