Multi-channel coherently driven reservoir and its operation method

Abstract

A multi-channel coherent drive reservoir and a method for operating the multi-channel coherent drive reservoir are provided. Embodiments of this disclosure utilize multiple physically separate delay loops coupled together using one or more couplers. Each loop provides a different connectivity scheme between virtual time nodes of the generated network.

Description

Technical Field

[0001] This disclosure relates to the field of optical computing, and more particularly to a multi-channel coherently driven reservoir and its operation method. Background Technology

[0002] Reservoir computing (RC) is a recently developed, bio-inspired computational framework that employs fixed, highly nonlinear dynamical systems (called reservoirs) to dramatically increase the dimensionality of sequential data. RC originates from recurrent neural network theory and is considered an intermediate approach combining simple training with high-dimensional nonlinear mappings based on physically realizable dynamical systems.

[0003] The main categories of photonic reservoir computing platforms utilize delay feedback architectures. Specifically, a novel all-optical delay dynamics system was recently demonstrated. In this system, the core of the photonic reservoir comprises a coherently driven fiber cavity, rather than directly integrating optoelectronic components typically used in delay feedback loops into the photonic reservoir. The absence of optoelectronic modules may improve the speed of information processing.

[0004] Furthermore, coherently driven fiber optic cavities do not require amplifiers or other active components. Since the absence of amplifiers or other active components eliminates the main noise sources, system performance is improved. Therefore, systems with coherently driven fiber optic loop cavities can meet the requirements of ultra-high-speed optical communication and information processing.

[0005] However, coherently driven fiber loops have a relatively simple interconnect network, while RC operations lie on high-dimensional mappings within highly interconnected echo-state networks. Therefore, the computational performance of coherently driven fiber loop reservoirs degrades in highly nonlinear tasks. Consequently, systems with coherently driven fiber loop reservoirs may be unable to complete tasks requiring higher-dimensional mappings due to their limited computational power.

[0006] The above information is provided to disclose information that may be relevant to this disclosure. It is not intended to acknowledge, nor should it be construed as, any of the above information constituting prior art relative to this disclosure. Summary of the Invention

[0007] This disclosure provides a method for operating a reservoir of multi-channel coherent drives and a reservoir of multi-channel coherent drives. According to this disclosure, a reservoir of multi-channel coherent drives is provided, comprising: a common mask function generator for generating at least one mask input signal using at least one input signal, the at least one input signal being periodically kept constant within the period of the common time mask function. The reservoir further comprises: a plurality of physically separated side-coupled loops, each loop including a delay line for delaying the at least one mask input signal, the delay line being desynchronized in a different manner with other delay lines associated with the side-coupled loop for creating a more complex interconnect network architecture. The reservoir further comprises: a plurality of directional couplers, each associated with one of the side-coupled loops, at least one of the directional couplers being used to inject the at least one mask input signal into one of the side-coupled loops; and a plurality of photodetectors, each photodetector associated with one of the side-coupled loops and used to detect one or more virtual time nodes created on the loops. The technical advantages of the reservoir of multi-channel coherent drives may include higher computing power. The increased number of time nodes and the more complex network architecture allow for enhanced computational power. For example, the feature of multiple side-coupled loops being desynchronized in different ways relative to the time mask can create networks with more complex interconnect structures, resulting in higher computational capabilities. Another technical advantage of multi-channel coherently driven reservoirs is the parallelization of the readout phase. Specifically, multiple readout channels can largely overcome the speed limitations of the detection phase. Another technical advantage of multi-channel coherently driven reservoirs can include suitability for on-chip implementation. For example, components such as integrated tunable delay lines, tunable couplers, and add-drop filters can be integrated onto a silicon chip.

[0008] In some embodiments, the delay time associated with the delay line is determined based on the mask period, the number of virtual time nodes, and the offset associated with the delay line.

[0009] In some embodiments, the reservoir further includes a tunable directional coupler (TDC) for coupling the side-coupled loops to each other. In some embodiments, the TDC is also used to couple the at least one masked input signal delayed by the delay line to an adjacent loop. Using the TDC, the network can be directly tuned and optimized (i.e., the network's direct tunability). This can be particularly advantageous for tasks where the nonlinear mapping can be directly adjusted.

[0010] As described above, the technical advantage of multi-channel coherently driven reservoirs can be their suitability for on-chip implementation (e.g., integrating components within a silicon chip). An integrated coherently driven reservoir with one or more tunable on-chip components can be faster than a reservoir computing (RC) device made from off-the-shelf components. Furthermore, integrated coherently driven reservoirs can be low-cost and scalable. Additionally, coherently driven reservoirs can exhibit sensitivity to environmental conditions, including temperature variations (e.g., ambient temperature variations). In some embodiments, one or more of the common mask function generator, the plurality of side-coupled loops, the plurality of directional couplers, and the plurality of photodetectors are integrated within the silicon chip. In some embodiments, the delay line and a plurality of tunable Bessel filters associated with the side-coupled loops for desynchronizing the side-coupled loops are integrated within the silicon chip. In some embodiments, the delay line is a helical waveguide. In some embodiments, each tunable Bessel filter includes a coupled ring optical waveguide (CROW), which is controlled using one or more integrated heaters. In some embodiments, the reservoir further includes a micro ring resonator (MRR) based add-drop filter for coupling the side-coupled loops to each other.

[0011] According to embodiments of this disclosure, a method for operating a reservoir of multi-channel coherent drives is provided. The method includes: receiving at least one input signal; periodically maintaining the at least one input signal constant within a period of a common time masking function. The method further includes: generating at least one masked input signal using the at least one input signal and the common time masking function; injecting the at least one masked input signal into the reservoir, the reservoir including a plurality of side-coupled loops, each of the side-coupled loops being physically separated from each other. The method further includes: delaying the at least one injected input signal using delay lines associated with each of the side-coupled loops, each of the delay lines being desynchronized with each other in a different manner to create a more complex interconnect network architecture. The method further includes: for each of the side-coupled loops, detecting one or more virtual time nodes created on the loop using a photodetector associated with the loop. The method further includes: outputting a response based on the detected virtual time nodes.

[0012] In some embodiments, the reservoir receives an input signal, and injecting the at least one masked input signal comprises: injecting the masked input signal into a first loop of the side-coupled loop via a directional coupler associated with the first loop. In some embodiments, delaying at least one injected input signal comprises: delaying the injected input signal by propagation on a first delay line; coupling the delayed input signal to a next loop via a tunable directional coupler (TDC), the next loop being adjacent to the loop associated with the first delay line; and delaying the coupled input signal on a second delay line associated with the next loop.

[0013] In some embodiments, the reservoir simultaneously receives two or more input signals, and injecting the one or more masked input signals comprises: injecting each masked input signal into the corresponding loop of the side-coupled loop via a directional coupler associated with the corresponding loop. In some embodiments, delaying at least one injected input signal comprises: delaying each injected input signal by propagation on a delay line associated with the corresponding loop; and coupling each delayed input signal to a next loop adjacent to the corresponding loop via a tunable directional coupler (TDC).

[0014] In some embodiments, the delay time associated with each delay line is determined based on the mask period, the number of virtual time nodes, and the offset associated with the delay line.

[0015] The technical advantages of the method for operating the above-described multi-channel coherently driven reservoir can substantially correspond to one or more of the technical advantages of the multi-channel coherently driven reservoir as defined above.

[0016] This disclosure also provides a computer program and a computer-readable medium for operating a reservoir of multi-channel coherent drive. The computer program and the computer-readable medium correspond to the methods and features of the reservoir described above for operating a reservoir of multi-channel coherent drive.

[0017] Embodiments have been described above in conjunction with various aspects of this disclosure, and these embodiments can be implemented based on these aspects. Those skilled in the art will understand that embodiments can be implemented in conjunction with the aspects describing these embodiments, but may also be implemented together with other embodiments of that aspect. It will be apparent to those skilled in the art that embodiments are mutually exclusive or incompatible with each other. Some embodiments may be described in conjunction with one aspect, but may also be applicable to other aspects, as will be apparent to those skilled in the art. Attached Figure Description

[0018] Furthermore, other features and advantages of this disclosure will be readily apparent from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0019] Figure 1 This is a diagram illustrating the basic concepts of feedforward neural networks, recurrent neural networks, and reservoir computing (RC) provided by existing technologies.

[0020] Figure 2 This is a diagram illustrating coherently driven and non-coherently driven reservoirs in the delay feedback architecture provided by existing technologies.

[0021] Figure 3A and Figure 3B This is a diagram illustrating the cyclic nonlinear dynamics in existing coherently driven reservoirs and equivalent cyclic networks provided by the prior art.

[0022] Figure 4 It is a diagram provided by existing technology showing time nodes at two different time steps.

[0023] Figure 5 This is a diagram of existing frequency multiplexing parallel RC that performs high-dimensional mapping, provided by existing technology.

[0024] Figure 6A and Figure 6B This is a diagram illustrating the operating principle of the dual-channel coherently driven reservoir provided in this embodiment of the disclosure.

[0025] Figure 7 This is a diagram illustrating a multi-channel coherently driven reservoir provided in an embodiment of this disclosure.

[0026] Figure 8 This is a diagram illustrating the input parallelization scheme in the reservoir of a multi-channel coherent drive provided in this embodiment of the disclosure.

[0027] Figure 9 This is a diagram illustrating an RC system using a side-coupled fiber optic ring to implement multi-channel coherent driving, as provided in an embodiment of this disclosure.

[0028] Figure 10 This is a diagram illustrating the hardware implementation of the on-chip dual-channel coherent drive RC provided in the embodiments of this disclosure.

[0029] Figure 11A This is a diagram illustrating an example of a coupled ring optical waveguide (CROW) filter integrated in an on-chip dual-channel coherently driven RC according to an embodiment of this disclosure.

[0030] Figure 11BThis is a diagram illustrating the characteristics of a CROW filter integrated in an on-chip dual-channel coherent drive RC, as provided in an embodiment of this disclosure.

[0031] Figure 12 This is a diagram illustrating the enhanced performance of embodiments of the present disclosure for performing high-order nonlinear tasks.

[0032] Figure 13 This is a flowchart of a method for operating a multi-channel coherent drive reservoir provided in an embodiment of this disclosure.

[0033] It should be noted that the same features are identified by the same reference numerals throughout the accompanying drawings. Detailed Implementation

[0034] This disclosure provides a reservoir for multi-channel coherent driving and a method for operating the reservoir for multi-channel coherent driving. Embodiments of this disclosure utilize multiple physically separate delay loops coupled together using one or more couplers. Each loop provides a different connectivity scheme between virtual time nodes of the generated network. Therefore, the coupled loops as a whole can realize a highly complex echo-state network with a tunable network connectivity scheme. Such a network can exhibit superior computational performance.

[0035] According to embodiments, delay feedback RC operates based on time multiplexing that effectively acts as a clock. Virtual time nodes are created by periodically multiplexing a (common) time mask of data sequences (synchronized with the data sequences). In various embodiments, coherently driven RC operates based on desynchronization of the mask and delay times (loops), and different time offsets can create different network architectures. Coherently driven RC employs multiple coupled delay loops, which are desynchronized in different ways relative to the mask.

[0036] Feedforward neural networks are a type of artificial neural network in which connections between nodes do not form loops. While feedforward neural networks are valuable for recognizing non-sequential data and are easy to train, they often suffer from memory shortages and cannot predict sequential data that requires memory.

[0037] The descendant of feedforward neural networks is the recurrent neural network (RNN). RNNs are also valuable for recognizing non-sequential data. RNNs have sufficient memory to predict sequential data that requires memory, thus overcoming the memory problem of feedforward neural networks. However, training RNNs is not simple and can be challenging, with high computational costs.

[0038] Derived from recurrent neural network theory, reservoir computing (RC) is a recently developed, bio-inspired computational framework that employs a fixed, highly nonlinear dynamic system (called a reservoir) to greatly increase the dimensionality of sequential data.

[0039] Recurrent Probability (RC) is considered an intermediate approach combining simple training with high-dimensional nonlinear mapping based on physically realizable dynamical systems. Essentially, RC maps sequential input data to a higher-dimensional computational space through the dynamics of a reservoir. One advantage of RC over recurrent neural networks is that training occurs only during the readout phase. Once the input signal is injected into a fixed reservoir, a readout mechanism is trained to read the reservoir's state and map it to the target output.

[0040] The basic concepts of feedforward neural network 101, recurrent neural network 102, and reservoir calculation 103 are as follows: Figure 1 As shown in the image.

[0041] Relative Logic (RC) can be extended to networks exhibiting memory decay (echo-state networks). RC is not limited to digital networks but can be extended to physical systems that are not networks in the classical sense but rather continuous dynamical systems. Major categories of physical RC are implemented in the photonic domain. Photonic RC opens up possibilities for real-time computing and can significantly enhance the computational power of machines that process information in real time, thus paving the way for next-generation data communications. Photonic RC utilizes high-dimensional optical / optoelectronic systems for nonlinear mapping.

[0042] Photonic RC systems can be designed to operate near chaotic boundaries (e.g., opto-chaos). This can provide hypersensitive responses to slight changes in input excitation, low latency, and low power budget. Hypersensitive responses can further lead to strong high-dimensional mappings, which may require a system with sufficiently high degrees of freedom, cyclic loops, and memory decay. Low latency can be achieved, in particular, through all-optical processing features. Furthermore, such photonic systems offer the potential for chip integration or incorporation into larger neuromorphic platforms.

[0043] Optical systems that meet the requirements of ultrasensitive response to input changes, low latency, and low power budget include nonlinear coupled-cavity systems, coupled chaotic lasers, and delay-feedback systems. Among these systems, delay-feedback systems are considered the most successful reservoir implementations due to their high dimensionality, ease of implementation, and tunability. The main categories of photonic RC platforms utilize delay-feedback architectures. This architecture typically relies on optoelectronic modulators and detectors within the delay-feedback loop. Therefore, the optoelectronic-optical conversion imposes fundamental limitations on such optoelectronic modulators and detectors in terms of speed and computational efficiency.

[0044] However, recently, a novel all-optical delay dynamics system has been demonstrated. In this new system, the delay feedback architecture includes a photonic reservoir, specifically a coherently driven fiber loop reservoir. Specifically, the core of the photonic reservoir comprises a coherently driven fiber cavity, rather than directly integrating optoelectronic components into the reservoir. It's important to note that in the coherent reservoir, the optical phase within the reservoir remains undisturbed, and the reservoir's response is essentially instantaneous; the disturbance is part of the RC dynamics.

[0045] The operation of a system with a coherently driven fiber optic loop cavity relies on the instantaneous coherent dynamics of light in an undisturbed optical system. The absence of optoelectronic modules can potentially increase the speed of information processing. Furthermore, the coherently driven fiber optic loop reservoir is a passive optical cavity, resulting in low loss. Therefore, amplifiers or other active components are not required within the cavity. Since the absence of amplifiers or other active components eliminates noise sources, system performance can be improved.

[0046] Therefore, systems with coherently driven fiber optic loop cavities can meet the requirements of ultra-high-speed optical communication and information processing. Coherently driven fiber optic loop cavities combine the advantages of methods based on delay dynamics systems and methods based on linear optical circuits with second-order nonlinearity provided by readout photodiodes.

[0047] On the other hand, in an incoherent reservoir, there is usually a nonlinear node that processes the amplitude of light rather than its phase, and the node’s response also acts as a low-pass filter. Figure 2 The diagram illustrates a coherently driven reservoir 210 and a non-coherent reservoir 220 in a delayed feedback architecture. The coherently driven reservoir 210 consists of Pauwels, The incoherent reservoir 220 was proposed by Vinckier, Quentin, et al. in "Photonic coherent reservoir computer based on fiber-ring with distributed nonlinearity" (Semiconductor Lasers and LaserDynamics IX, Vol. 11356, IEEE Optical Engineering Society, 2020). The incoherent reservoir 220 was proposed by Vinckier, Quentin, et al. in "High-performance photonic reservoir computer based on a coherently driven passivecavity" (Optica 2.5 (2015): 438-446).

[0048] A key aspect of understanding the working principle of coherently driven fiber loop reservoirs is time multiplexing. For example, it provides... Figure 3A and Figure 3B This illustrates the cyclic nonlinear dynamics 310 and the equivalent cyclic network 320 in a conventional coherently driven reservoir. Figure 3A In the cyclic nonlinear dynamics 310, a sample-and-hold process is used to periodically keep the amplitude-encoded input signal constant for a duration T0. This input sequence (e.g., a sequence of input symbols) is then multiplied by a random masking function. For example, the masking function is a periodic time masking function with a time step of θ and a period of T0. The masked signal is then injected into a reservoir, creating a [missing information - likely a specific type of mask]. The reuse time corresponds to N virtual time nodes. This reserve pool has a delay feedback architecture with a delay time of T. delay .

[0049] In a coherently driven fiber optic loop, the interconnection between internal variables is achieved by synchronizing the mask input and the delay time. If the periodic mask (acting as the "clock" of the RC circuit) and the delay time T... delay With T delay Precise desynchronization of the time offset –T0=k×θ (where k is an integer) enables rich cyclic dynamics. In this scenario, a virtual time node with an index i (i=1, 2, ..., N) at a given time step is connected to the node i–k at the previous time step.

[0050] Figure 3B The equivalent recurrent network 320 illustrates this feature. (Reference) Figure 3B Black node 321 is a virtual time node created through time multiplexing. If T... delay –T0=k×θ, then the equivalent cyclic network 320 is similar to two rings offset by k nodes, such as Figure 3B As shown.

[0051] Furthermore, to provide complex connectivity architectures, the latency of the loop (e.g., a coherently driven fiber loop) differs from the timing of the input data and the period (T0) of the mask, as shown in the following equation:

[0052]

[0053] On the other hand, in traditional optoelectronic reservoirs, the delay time associated with the feedback loop is essentially equal to the period T0 of the masking function. Therefore, this leads to synchronization of the masking and feedback times. Thus, as... Figure 4 As shown, each time node with a time step of n is coupled to the same time node with a time step of n–1.

[0054] A system with a coherently driven fiber optic loop reservoir can be used as an RC platform. Despite the lack of nonlinear nodes within the cavity, the coherently driven reservoir shown above (e.g.) Figure 2 The coherent reservoir 210 in the image exhibits superior performance in simpler machine learning tasks that require cyclic dynamics. In fact, coherently driven fiber loop reservoirs can exhibit excellent performance in several processes (e.g., processing at very high speeds), such as information processing speeds up to 50 GHz or even higher.

[0055] However, coherently driven fiber loop reservoirs have a fairly simple interconnect network, while RC operations lie on a high-dimensional mapping within a highly interconnected echo-state network. Therefore, performance differs for higher-order nonlinear tasks. In fact, the performance of coherently driven reservoirs degrades for higher-order nonlinear tasks. For example, the performance of coherently driven reservoirs (e.g., the computational performance of coherently driven fiber loop reservoirs) may be reduced for highly nonlinear tasks because the equivalent network associated with the coherently driven reservoir has a fairly simple interconnect structure, and the nonlinearity of the coherently driven reservoir is limited to the photodetector outside the cavity during the readout phase. Therefore, systems with coherently driven fiber loop reservoirs may not be able to complete tasks requiring higher-dimensional mappings due to lower computational power. Furthermore, at least in some cases, the sequential detection of time nodes via the photodetector also negatively impacts reservoir performance. In other words, the maximum time node detection speed further limits the number of usable time nodes, thus negatively affecting the computational performance and capacity of the reservoir.

[0056] To overcome the aforementioned problems, various parallelization methods have been envisioned. Parallel RC is a method of de-embedding nodes in time, space, or frequency. For example, parallel photonic reservoir computers have been provided, in which neurons (or virtual nodes) are implemented as physically separated delay lines on a silicon chip. Although such photonic reservoir computers include spatially multiplexed nodes, effectively eliminating the need for fast time multiplexing, some drawbacks have been found. For example, the connections of spatially multiplexed nodes in a parallel photonic reservoir computer are static and strictly dependent on the topology of the silicon chip. Assuming complex tasks require a large number of nodes, static and topology-dependent connections may require relatively large chip sizes and large photodetector arrays. Furthermore, compared to tasks performed by delay reservoirs, some parallel RC systems (such as the RC shown by Vandoorne, Kristof et al. in "Experimental demonstration of reservoir computing on a silicon photonics chip", Nature Communications 5.1 (2014):1-6) have only been tested on fairly basic tasks. The full capacity of parallel optical RC is still unknown; therefore, proving the full capacity of parallel optical RC remains one of the unresolved challenges in this field.

[0057] Parallel RC can open a new avenue for real-time, high-speed data processing. Coherently driven reservoirs can be parallelized to keep pace with high-speed data communication. Most successful parallel RC work employs multiple frequency channels created by embedded phase modulators. The responses of these systems are instantaneous and coherent. However, parallelism can have some identifiable drawbacks. For example, resolving frequency channels during the readout phase can be challenging, the dynamics are fixed and untunable, and task-specific settings are limited.

[0058] Figure 5 The parallel RC circuit with frequency multiplexing is shown. Figure 5The frequency-multiplexed parallel RC 500 shown is proposed by Butschek, Lorenz, et al. in “Parallelphotonic reservoir computing based on frequency multiplexing of neurons” (arXiv e-prints(2020):arXiv-2008). Using frequency multiplexing, the frequency-multiplexed parallel RC 500 includes a delay loop 510 that performs high-dimensional mapping. The coherently driven loop 510 can create multiple frequency channels using a phase modulator 520. The state of the reservoir 500 is encoded in the amplitude and phase of the frequency sidebands of a highly coherent laser 540 that travels or propagates in a single-mode, polarization-preserving fiber loop 511. Nodes that create the reservoir in the frequency and time domains, and the created frequency channels, are coupled by a single nonlinear node implemented by the phase modulator 520. However, the response of the coherently driven reservoir 500 is instantaneous, and the dynamics of the reservoir 500 can be further enhanced, for example, through desynchronization via masking and delay time. Furthermore, the number of frequency nodes determined by the incorporated phase modulator 520 based on its driving speed is limited, and therefore the performance of the RC 500 may be limited due to an electrical bottleneck. Additionally, while the dynamics of the reservoir 500 are quite rich, the detection of time-frequency nodes may require complex demultiplexing modules (e.g., DeMux 530). Moreover, resolving the frequency channels during the readout phase can be challenging (complex readout), and a large number of nodes requires wideband and dispersion compensation settings. Furthermore, parallel frequency channels are not directly accessible.

[0059] Current coherently driven RC platforms offer several advantages over optoelectronic RC platforms, such as faster operating speeds and lower noise. Furthermore, coherently driven RC platforms are less constrained by electrical bottlenecks. However, as mentioned above, current coherently driven RC platforms typically have lower computational power for tasks requiring higher-dimensional mappings. For example, the equivalent discrete network associated with a current coherently driven reservoir has a fairly simple interconnected loop structure. Each node (e.g., Figure 3B Node 321 in the pool is connected to only a single node in the preceding time steps (i.e., connected to only one node). This can limit computational power (e.g., exhibiting low computational performance), especially when the pool performs highly complex tasks (e.g., highly nonlinear tasks).

[0060] Furthermore, assuming nodes are detected by photodetectors, the number of virtual nodes (i.e., N) is limited by the integration time of the photodetectors. In other words, the maximum number of virtual nodes can be limited by the sequential detection scheme of the readout phase. This limitation may be due to the fact that node detection occurs serially (e.g., node by node), and the detection speed limits the number of nodes that can be used (θ ~ τ). PD Furthermore, the limited bandwidth of the photodetector determines the fastest timescale for creating virtual nodes. The number of virtual nodes (i.e., N) limited by the integration time of the photodetector can be defined as follows:

[0061]

[0062] Reservoir computing can be efficiently used for high-speed real-time information processing and data classification (e.g., real-time nonlinear channel equalization, nonlinear distortion compensation for power amplifiers in wireless communication, Internet Protocol (IP) packet classification, speech recognition, sensing, and molecular recognition). Reservoir computing (RC) can be an excellent analog computing platform, highly advantageous for compensating for nonlinear distortion in optics and wireless communication. RC can complement another class of neuromorphic optical computing platforms, enabling real-time data processing. For example, cascaded reservoirs and feedforward neural networks (FNNs) can form powerful real-time processors. However, integrating RC on silicon chips is likely a major challenge in silicon photonics. Successful implementation relies on large off-the-shelf components (e.g., large photodetector arrays).

[0063] While chaotic optical systems can be considered an alternative to state-of-the-art delay feedback (coherent or incoherent) platforms, their operating speed may be limited by the detection bandwidth during the readout phase. Other parallelization alternatives are still conceivable.

[0064] This disclosure provides methods and apparatus for overcoming the aforementioned challenges and problems by providing multiple side-coupled coherently driven loops (e.g., physically separated loop cavities or channels) and associated physically separated delay lines. The multiple side-coupled coherently driven loops are desynchronized in different ways relative to a time mask. This feature allows for the creation of networks with more complex interconnect structures, thereby providing greater computational power. Each channel (i.e., each loop or loop cavity) has its own photodetector, thus enabling the detection of more nodes in parallel. The multiple side-coupled coherently driven loops can be coupled via tunable couplers, so the equivalent network can be designed and tuned according to the task.

[0065] According to embodiments, a method is provided to substantially maximize or optimize the operating speed and computing power of a photon reservoir computer. The architectures in various embodiments utilize multiple delay lines to increase the number of virtual time nodes initially created by external time multiplexing. In various embodiments, the delay lines are desynchronized relative to a time mask function. Each delay line can be desynchronized in a different manner (e.g., each delay line is desynchronized in a manner different from the other delay lines). This configuration can create complex interconnect architectures. According to embodiments, computing power can be significantly increased due to the greater number of time nodes and the more complex network architecture.

[0066] In various embodiments, a tunable Bezier delay line is used to adjust the offset. In some embodiments, additional tunability can be achieved through a tunable coupling network that can adjust the network according to the complexity of the task to be performed.

[0067] According to embodiments, a photonic reservoir computing (or photonic reservoir computer) is provided, including multiple loop cavities (or channels) side-coupled via multiple tunable directional couplers. The loop cavities are used to create more complex cyclic interconnect structures in a coherently driven platform. In various embodiments, each loop or loop cavity acts as a delay feedback loop system, wherein the delay lines are desynchronized relative to a common periodic mask (e.g., a clock) at the input. If the loops or loop cavities have different delays relative to the common mask, the dynamics of the equivalent network will be greatly enhanced. Different desynchronization schemes result in different interconnect network architectures. Furthermore, each node can connect to multiple nodes in a previous time step.

[0068] In various embodiments, the more complex loop system not only enhances the dynamics but also allows for the spontaneous creation of multiple time nodes in parallel within the loop or loop cavity. Therefore, for a given time mask, more time nodes are generated in parallel.

[0069] Furthermore, assuming the loops or loop cavities are physically separated, the state of each loop or loop cavity can be detected by a separate photodetector. Therefore, the state of the reservoir can be detected in parallel by multiple photodetectors.

[0070] According to embodiments, different numbers of physically separated channels can be used. For some high-order nonlinear tasks that may require complex dynamics, increasing the number of channels (or loop cavities) with different node offsets (e.g., node desynchronization) can potentially improve performance. In various embodiments, the different channels (or loop cavities) can be directly accessed via a photodetector (PD) associated with each channel.

[0071] According to embodiments, different network architectures can be designed by adjusting the delay time. The architecture of the equivalent network can be directly adjusted according to the complexity of the task. Dynamics can be directly adjusted by selecting different node displacement schemes through a tunable delay line. In various embodiments, a suitable type of network model for the task is initially identified, and then the network is implemented by adjusting the relevant delays (e.g., by adjusting a delay knob). In some embodiments, a tunable integrated platform provides the possibility of designing the network by controlling a knob (or delay knob). In some embodiments, further adjustments (or tuning) are provided through the intra-channel coupling coefficients. The intra-channel coupling coefficients can be adjusted by directional couplers embedded in the network, thus allowing for sufficient optimization of intra-channel coupling.

[0072] According to the embodiments, the network provides high computational power (i.e., very rich computational performance) by combining the effects of a larger number of time nodes and the complex connections caused by different desynchronization schemes. This creates a large number of time nodes using multiple physically separated channels, which can eliminate the need for ultrafast detectors by detecting nodes in parallel.

[0073] According to an embodiment, the inputs can be parallelized in a multi-channel coherently driven reservoir. For example, multiple inputs can be injected in parallel into a multi-channel RC circuit. This input parallelization is particularly advantageous for wireless communications that present signals on an in-phase and quadrature (IQ) basis.

[0074] Multi-channel parallel coherent driven RC

[0075] This disclosure provides a multi-channel parallel coherently driven RC, which can be considered a novel coherent reservoir computing platform that leverages spatial parallelism and complexity to enhance computational power. According to an embodiment, multiple time nodes are created in parallel across multiple physically separated delay lines. The multiple time nodes are initially created via external time multiplexing. Because these delay lines are desynchronized differently relative to a common mask (clock), each delay line provides a unique connectivity scheme (e.g., a complex interconnect architecture). Desynchronization of the different delay lines can be tuned using integrated tunable delay lines.

[0076] The operating principle of a multi-channel coherently driven reservoir is shown below. According to an embodiment, optimizing the reservoir's performance may require adjusting its configuration to achieve the optimal balance between in-channel coupling and loop gain.

[0077] Figure 6A and Figure 6B The operating principle of the dual-channel coherently driven reservoir provided in this disclosure embodiment is illustrated. Specifically, Figure 6AFIG. 0 shows a dual-channel coherently-driven RC 610 with two spatially-separated delay lines 611 and 612 (i.e., DL1 and DL2). Delay line 611 and delay line 612 are coupled side-by-side by a tunable directional coupler (TDC) 613. The masked input signal 614 has a masking period of T0 and is injected into the first channel 615a (i.e., the first loop) through a directional coupler 616 (i.e., DC1). The delay times associated with delay line 611 and delay line 612 are respectively and The states of loop 615a and loop 615b are read by photodetectors 617 and 618 (i.e., PD1 and PD2), respectively.

[0078] The working principle of a multi-channel coherently-driven reservoir is further described below with reference to Figure 6A DEFINITION. It should be noted that, for simplicity, the simplest implementation of a multi-channel coherently-driven reservoir (e.g., the dual-channel coherently-driven RC 610) is used to illustrate the concept of the operating principle. However, those skilled in the art will readily understand that more than two channels (i.e., multiple side-coupled loops) can be employed in various embodiments of a multi-channel coherently-driven reservoir, as further shown in Figure 7 As further shown. It should also be noted that, like traditional coherently-driven fiber-loop reservoirs, time multiplexing is crucial in the operation of a multi-channel coherently-driven reservoir.

[0079] REFERENCE Figure 6A , initially, using a sample-and-hold procedure, the amplitude-encoded input signal is periodically held constant for a duration T0. Then, the input signal is multiplied by a masking function, which is periodic with a period of T0. The masking function has multiple random levels and divides each time interval T_0 into N time nodes The masked input signal 614 is (optically) injected into the reservoir 610 through a directional coupler 616.

[0080] Subsequently, the injected light travels (e.g., propagates) on delay line 611 and creates N virtual time nodes. The injected signal is delayed by where k1 < N is an integer. The delayed signal is coupled to the second loop 615b through a tunable directional coupler (TDC) 613. The coupling coefficient can be adjusted to obtain the desired performance of the RC. The coupled signal (e.g., the signal coupled to the second loop 615b) is again delayed on the second delay line 612. The second delay time is where k2 < N is an integer.

[0081] The time nodes created in each loop (e.g., N nodes per loop) are detected by photodetectors 617 and 618 (i.e., PD1 and PD2), respectively. Finally, the states of the time nodes in the reservoir are used for training and classification. There are a total of 2×N time nodes because there are two channels in this example (e.g., dual-channel coherently driven RC 610).

[0082] Figure 6B An equivalent cyclic network 620 describing the discrete dynamics of a dual-channel coherent RC 610, as provided in an embodiment of this disclosure, is illustrated. Each time node at time step n×T0 is cyclically connected to two independent nodes at time step (n–1)×T0. In various embodiments, the cyclic interconnection structure between two successive time steps can be adjusted by selecting one or more of integers k1 and k2.

[0083] Integers k1 and k2 determine the time offsets of loops (e.g., loops 615a and 615b) such that a node with index i at time step n×T0 can be coupled at loop 615a and loop 615b to nodes with indices i–k1 and i–k2 at time step (n–1)×T0, respectively. Integers k1 and k2 are typically small integers and can be distinct numbers. The offsets defined by integers k1 and k2 determine the network interconnect architecture and therefore play a crucial role in the RC performance of high-dimensional mappings in the computational domain. Typical choices for offset integers k1 and k2 are in the range of 1 to 4, which maintains network sparsity but is rich in nonlinear cyclic dynamics.

[0084] Figure 7 A multi-channel coherent drive reservoir 700 according to an embodiment of this disclosure is shown. As described above, those skilled in the art will readily understand that the operating principle of a dual-channel coherent drive reservoir can be extended to larger cavity systems including multiple side-coupled coherent drive loops (e.g., Figure 7 The system is a multi-channel coherently driven reservoir (700). It is believed that this system can provide a very rich set of nonlinear dynamics.

[0085] refer to Figure 7 The multi-channel coherently driven reservoir 700 includes M side-coupled loops 711, 712, 713, ..., 71M. Multiple side-coupled loops 711, 712, 713, ..., 71M (where i = 1, 2, ..., M) with spatially separated delay lines 721, 722, 723, ..., 72M are coupled via tunable directional couplers (TDCs) 731, 732, 733, ..., 73(M–1).

[0086] Delay lines 721, 722, 723, ..., 72M are desynchronized relative to the common input mask, so that delay lines 72i (where i = 1, 2, ..., M) provide delay time. Where, k i It is an integer 0 <k i <N。

[0087] Tunable directional couplers (TDCs) 731, 732, 733, ..., 73(M–1) (i.e., TDC 73i, where i = 1, 2, ..., M–1) are used to side-couple loops (e.g., loop cavities or channels) 711, 712, 713, ..., 71M with a tunable coupling ratio that provides additional control over the dynamics of the reservoir 700. Through tunable coupling, the network can be further tuned to provide additional paths for controlling the nonlinear dynamics of the reservoir 700. Each loop includes a delay line desynchronized relative to a common input mask.

[0088] The states of loops 711, 712, 713, ..., 71M are measured by individual photodetectors 741, 742, 743, ..., 74M (i.e., PD 74i, where i = 1, 2, ..., M). Directional couplers (DC) 751, 752, 753, ..., 75M (i.e., DC 75i, where i = 1, 2, ..., M) are used to couple the optical output and measure the state of each loop.

[0089] The multi-channel coherently driven reservoir shown above can offer many benefits. According to an embodiment, the multi-channel coherently driven reservoir can have higher computational power. When there are N time nodes and D delay lines, D×N nodes will be created. The more nodes, the higher the computational power.

[0090] Unlike traditional coherently driven single-loop reservoirs that are cyclically connected to only one node at each time point, the multi-channel coherently driven reservoir of this application implements a more complex interconnection network structure. Therefore, the multi-channel coherently driven reservoir of this application can provide a richer variety of cyclic nonlinear dynamics. In other words, the connectivity of the multi-channel coherently driven reservoir is much more complex than that of the single-channel coherent reservoir, resulting in a much richer variety of nonlinear transformations.

[0091] One advantage of this greatly enhanced nonlinear dynamics is that it increases the computational power of the reservoir, especially when dealing with tasks that require strongly nonlinear mappings in the computational space.

[0092] Furthermore, tunable directional couplers (i.e., the direct tunability of the network) can be used to directly tune and optimize the network. This is particularly advantageous for tasks where the nonlinear mapping should be directly tuned. The network architecture can also be tuned by adjusting the delay within the time scale θ (using tunable delay lines on each channel).

[0093] Other benefits of multi-channel coherently driven reservoirs include the advantage of parallelization during the readout phase. Specifically, multiple readout channels can largely overcome the speed limitations of the detection phase. According to an embodiment, each channel has its own photodetector, so the state of each loop can be detected individually or separately by the channel's own photodetector. Since the loop state is detected individually by the photodetectors, the speed limitations associated with the detection phase can be largely avoided, as each photodetector needs to detect a smaller number of time points.

[0094] In contrast, in conventional delayed-feedback RC circuits, the maximum number of detectable nodes is typically limited by the bandwidth of the photodetector. For a given data rate, this can limit the number of nodes. According to the embodiments, by employing a separate photodetector for each loop, more time nodes can be detected without violating electrical bottlenecks.

[0095] Another benefit of multi-channel coherently driven reservoirs is their suitability for on-chip implementation. In various embodiments, components such as integrated tunable delay lines, tunable couplers, and add / drop filters can be efficiently integrated onto a silicon chip.

[0096] Input Parallelization

[0097] According to an embodiment, input parallelization can be implemented in a multi-channel coherently driven reservoir. The multi-channel coherently driven reservoir can simultaneously accept multiple inputs, which are multiplied by a common time mask. This input parallelization can be used for one or more ultra-high-speed real-time data processing and quadrature modulation schemes in wireless and optical communication links.

[0098] Input parallelization means injecting multiple input signals into the reservoir simultaneously and processing the information in parallel. In various embodiments of this disclosure, a multi-channel coherent reservoir can simultaneously accept multiple input time signals on its spatially separated channels (e.g., loops or loop cavities). If the input signals are multiplied by a common time mask, the reservoir operates according to the "timing" of the mask and maintains the network interconnect structure based on the desynchronization of the loop and the mask.

[0099] Figure 8 An input parallelization scheme in a reservoir 800 for multi-channel coherent driving provided in an embodiment of this disclosure is illustrated. (See also...) Figure 8, several input signals (such as input signals 861, 862, 863, ……, 86M) are simultaneously injected into the reservoir 800 through directional couplers 851, 852, 853, ……, 85M (i.e., DC 85i, where i = 1, 2, ……, M) respectively.

[0100] Each loop in loops (i.e., channels) 811, 812, 813, ……, 81M is equipped with a corresponding directional coupler (e.g., directional couplers 851, 852, 853, ……, 85M). Each directional coupler acts as a linear node and can couple the corresponding input signal into the reservoir 800. Since the input signals 861, 862, 863, ……, 86M are simultaneously injected into the reservoir 800, the reservoir 800 can effectively solve the non - linear damage between individual input channels. The way of adopting input parallelization in a multi - channel coherent - driven reservoir is as follows.

[0101] First, using a sample - and - hold program, the amplitude - encoded input signals (e.g., input signals 861, 862, 863, ……, 86M, M input channels) are periodically held constant within a duration T0. Then, the input signals 861, 862, 863, ……, 86M (i.e., multi - input) are multiplied by a common masking function 870 (which can be generated by a common masking function generator), and the common masking function 870 is periodic with a period of T0. Similar to the single - input case (e.g., Figure 6A the two - channel coherent - driven reservoir 610 shown in ), the masking function 870 has multiple random levels and divides each time interval T0 into N time nodes.

[0102] The masked input signals 871, 872, 873, ……, 87M are injected into the reservoir 800 (optically) through directional couplers 851, 852, 853, ……, 85M respectively. The masked signals 871, 872, 873, ……, 87M are directly injected into loops 811, 812, 813, ……, 81M respectively.

[0102] Then, the injected light travels (e.g., propagates) on the delay lines associated with each of the loops 811, 812, 813, ……, 81M and creates N virtual time nodes on each of the loops 811, 812, 813, ……, 81M. The injected signal is delayed by where k i <N is an integer. The delayed signal is coupled to the adjacent second loop 812 through a tunable directional coupler (TDC) 831.

[0103] The time nodes created on each of loops 811, 812, 813, ..., 81M (i.e., N virtual time nodes for each loop in loops 811, 812, 813, ..., 81M) are detected by photodetectors 841, 842, 843, ..., 84M (i.e., PD 84i, where i = 1, 2, ..., M). Finally, the states of the time nodes in the reservoir are used for training and classification. There are a total of M×N time nodes, because there are M channels in this case (e.g., a multi-channel coherently driven RC800).

[0104] According to embodiments, input parallelization provides a way to achieve ultrafast signal processing because multiple signals (e.g., symbols) can be simultaneously injected into a multi-channel coherently driven reservoir without requiring ultrafast time multiplexing or fast detection schemes. Assuming the reservoir operation requires time multiplexing at a rate much faster than the input bandwidth, the higher input data rate cannot be directly handled by a conventional reservoir. Instead, multiple samples can be held and sent to the reservoir using a common time mask. When holding “M” samples, time multiplexing can be “M” times slower, and therefore more practical. For example, in a binary data link with a bit rate of 25 GB / s, when 50 time nodes are needed to operate the reservoir, the time multiplexing mask should be 50 times faster. However, with input parallelization, the operation of the reservoir (e.g., a multi-channel coherently driven reservoir) can be more efficient. When there are “M” channels, “M” bits can be held and sent to the multi-channel coherently driven reservoir simultaneously, making time multiplexing “M” times slower.

[0105] There are several practical use cases where communication links contain multiple independent signals with linear / nonlinear impairments. Examples of such systems include orthogonal polarization channels in fiber optic communications, WDM optical communication links, and orthogonal modulation channels in wireless communications. One of the most useful applications of RC is real-time nonlinear distortion cancellation in communication systems. The communication channel shown above can utilize multi-input coherent RC (e.g., a multi-channel coherently driven RC with input parallelization) so that independent channels can be processed simultaneously by the RC.

[0106] On-chip implementation

[0107] As described above, one advantage of multi-channel coherently driven reservoirs, according to the embodiments, is their suitability for on-chip implementation. Tunable integrated delay lines, integrated directional couplers, and tunable add-drop filters can be used to implement on-chip coherently driven reservoirs. While reservoirs have some common settings, several parameters of coherently driven reservoirs can be fine-tuned, especially for highly nonlinear tasks. Therefore, the challenge of on-chip implementation of multi-channel coherently driven reservoirs is designing a highly stable yet tunable hardware platform to fine-tune parameters such as timing, in-channel coupling, loop attenuation, and input coupling ratio.

[0108] In the implementation of a tunable multi-channel coherently driven reservoir, the RC system comprises multiple delay lines. The delay time associated with each delay line can be individually adjusted with the precision of a virtual time node. The delay lines are side-coupled via tunable directional couplers. The coupling strength is a control parameter that adjusts the network connectivity. Readouts from different frequency channels can be performed in parallel, thus each channel has its own photodetector. In other words, each channel has its own photodetector.

[0109] In the hardware implementation of the tunable multichannel coherently driven reservoir of this application, the RC includes a coupled ring optical waveguide (CROW) as an integrated delay line. In various embodiments, the CROW implements a Bessel filter, which is synthesized to substantially minimize group delay distortion (i.e., to flatten the group delay to the greatest extent possible). The CROW can be discretized, for example using heaters, to achieve a time delay of approximately one virtual time node by turning the resonators on and off. In various embodiments, the delay line is side-coupled via a directional coupler, which can be tunable.

[0110] Figure 9 This illustration shows an RC implementation of multi-channel coherent drive using a side-coupled fiber optic loop, as provided in an embodiment of this disclosure. (Reference) Figure 9 Each ring (e.g., loop, loop cavity, or channel, such as loops 911, 912, 913) includes a delay T d = Fixed delay lines of T0 (e.g., delay lines 921, 922, 923) and lines with delay steps Tunable integrated delay lines (TIDL) (e.g., TIDL 951, 952, 953).

[0111] Loops 911 and 912 are laterally coupled to each other via tunable directional coupler 931, and loops 912 and 913 are laterally coupled to each other via tunable directional coupler 932. Tunable directional couplers 931, 932, and 933 can be adjusted according to one or more tasks being performed. Each loop in 911, 912, and 913 includes its own photodetectors 941, 942, and 943, respectively. Photodetectors 941, 942, and 943 read (virtual) time nodes (e.g., parallel detection) created on each delay line in delay lines 921, 922, and 923, respectively.

[0112] Figure 9 The implementation can be mathematically defined as follows:

[0113]

[0114]

[0115] In the above equation, there are several coupling parameters, including: C, representing intra-channel coupling; α, representing loop attenuation; β, representing input gain. Additionally, δ... l This represents the return phase of the loop. Nonlinear parameters also exist. The model includes the distributed Kerr effect, and the effective fiber length is L. eff The Kerr coefficient is This represents the i-th time node at time step n in channel c.

[0116] Figure 10 This is a schematic diagram of the hardware implementation of an on-chip dual-channel coherent drive RC provided in an embodiment of this disclosure. In the hardware implementation of the on-chip dual-channel coherent drive RC, the integrated component is a tunable integrated component (e.g., a tunable integrated filter). According to an embodiment, the RC integrating multi-channel coherent drive includes multiple physically separated channels, which have equal time delays associated with the loop.

[0117] exist Figure 10 In the illustrated hardware implementation, the RC 1000 with integrated dual-channel coherent drive includes two physically separate channels (loops or loop cavities) 1011 and 1012, each with an equal time delay associated with loops 1011 and 1012. Specifically, two static delay lines 1021 and 1022 in each of the channels (loops) 1011 and 1012 provide equal time delays, respectively. The two channels 1011 and 1012 are coupled via a micro-ring resonator (MRR) based add-drop filter 1060. The RC 1000 with integrated dual-channel coherent drive also includes directional couplers 1051, 1052, 1053, and 1054. The directional couplers 1051, 1052, 1053, and 1054 can be tunable and are used to couple light into and out of the RC 1000, and also to control loop attenuation.

[0118] Channels (loops) 1011 and 1012 include static delay lines 1021 and 1022, respectively. Static delay lines 1021 and 1022 are helical waveguides. The group delay associated with the helical waveguide is equal to the mask period, T0. The additional delay for desynchronizing the loops (e.g., loops 1011 and 1012) relative to the mask is provided by coupled ring optical waveguide (CROW) Bessel filters (e.g., CROW Bessel filters 1031 and 1032). Therefore, the total group delay for each loop in loops 1011 and 1012 is T0 + T.CROW , among which, T CROW It is the tunable delay time associated with the CROW filter (i.e., filter 1031 or filter 1032).

[0119] Each loop in loops 1011 and 1012 is cascaded with a tunable filter (referred to as a Bessel filter, e.g., CROW reflective Bessel filters 1031 and 1032) that provides maximum flatness of the group delay. Channels (loops) 1011 and 1012 are desynchronized via CROW 1031 and 1032. In various embodiments, CROW filters 1031 and 1032 are used as reflective Bessel filters. CROW reflective Bessel filters 1031 and 1032 are used to adjust the desynchronization of each loop in loops 1011 and 1012 relative to a common time mask. An example of a CROW filter integrated into the RC of an on-chip dual-channel coherent drive according to embodiments of this disclosure is shown in... Figure 11A Further illustrated below. According to one embodiment, the CROW filters 1031 and 1032 integrated in the dual-channel coherent drive RC 1000 are characterized in... Figure 11B As shown in the image.

[0120] Each of the CROW filters 1031 and 1032 includes multiple side-coupled ring resonators. Specifically, the loop cavities of CROW 1031 and 1032 are coupled via side-coupled ring resonators. In various embodiments, the resonant wavelengths of the side-coupled ring resonators can be adjusted by integrated heaters 1031h and 1032h, respectively. Heaters 1031h and 1032h can cause the resonant frequency of each resonator (side-coupled ring resonator) to drift, thus adjusting the overall group delay of the reflected CROW 1031 and 1032. Therefore, each coupling can be adjusted by heater 1031h or 1032h.

[0121] Each CROW filter can be discretely tuned by heaters 1031h or 1032h, for example, by the following definition:

[0122]

[0123] Where N is the number of time points at the input, and T0 is the duration of the input bit (i.e., the mask period).

[0124] CROW filters 1031 and 1032 can be optimized such that the delay time can be discretely changed in steps θ = T0 / N by turning heaters 1031h and 1032h on and off, where N is the total number of time nodes in each loop. The input mask signal 1040 is coupled to the reservoir 1000 via a directional coupler 1051. In some embodiments, the directional coupler 1051 is tunable, allowing direct control of the input gain (i.e., the intensity of the light coupled into the RC). In-channel coupling can be achieved through... Figure 10 The add-drop ring resonator (MRR-based add-drop filter 1060) shown is tuned. Therefore, the amount of coupling between channels 1011 and 1012 is adjusted by an integrated heater controlling the resonant frequency of the control loop (MRR 1060). The state of the reservoir 1000 is detected by directional couplers 1052 and 1053. In some embodiments, couplers 1052 and 1053 are tunable to control and fine-tune loop attenuation according to the computational complexity of the task being performed.

[0125] According to embodiments, an integrated coherently driven RC with one or more tunable on-chip components can be faster than an RC made from off-the-shelf components. Furthermore, the integrated coherently driven RC can be low-cost and scalable. Integrated components such as CROW filters can be precisely designed according to the bandwidth of the time mask. In addition, the coherent reservoir can exhibit additional sensitivity to environmental conditions, including temperature variations (e.g., ambient temperature variations). The integrated platform is highly stable and can operate without any stabilizers.

[0126] In various embodiments of this disclosure, the architecture of the multi-channel coherently driven RC utilizes spatial parallelization, wherein multiple coupled spatial channels operate and perform nonlinear mappings simultaneously. According to embodiments, a method is provided to create rich dynamics by desynchronizing channels differently relative to a common time mask. In various embodiments, multiple channels can be used to increase the number of time nodes, thereby improving computational performance at least in part due to the increased number of nodes. Figure 12 Simulations illustrating the enhanced performance of embodiments of this disclosure for performing high-order nonlinear tasks are shown. Each node on each channel (during the readout phase) uses a separate photodetector for detection, thus the RC benefits from parallelization at the readout. This eliminates the need for ultrafast electronic switches to create time masks and enables operation with a smaller number of nodes, for example, through node parallelization, while maintaining its computational power. In some embodiments, additional tunability is achieved through adjustable intra-channel coupling ratios or intra-channel coupling coefficients. This can be optimized according to the computational task. Embodiments of this disclosure also provide input parallelization. Multiple inputs (e.g., those with some nonlinear impairments) can be processed simultaneously.

[0127] Among other things, embodiments of this disclosure can be used for efficient information processing in a wide range of applications, from fast computation to optical communications, such as compensating for nonlinear channel distortion caused by Kerr nonlinearity, multipath interference, and nonlinear channel dispersion. Since the reservoir is fixed and training is performed only during the readout phase, reservoir computation can be processed without additional kinetic adjustments.

[0128] According to some embodiments, multi-channel coherently driven RC can be effectively deployed to utilize wavelength division multiplexing (WDM) for parallel information processing. Since the core of the coherently driven reservoir has no nonlinear nodes, the WDM channels can be processed in parallel, thereby achieving high-speed real-time information processing.

[0129] Figure 13 A method for operating a reservoir of multi-channel coherent drives, provided by an embodiment of this disclosure, is illustrated. The method includes: receiving (1310) at least one input signal; periodically maintaining (1320) at least one input signal constant within a period of a common time masking function. The method further includes: generating (1330) at least one masked input signal using the at least one input signal and the common time masking function; injecting (1340) the at least one masked input signal into a reservoir comprising a plurality of side-coupled loops, each of the side-coupled loops being physically separated from each other. The method further includes: delaying (1350) at least one injected input signal using delay lines associated with each of the side-coupled loops, each delay line being desynchronized with each other in a different manner to create a more complex interconnect network architecture. The method further includes, for each side-coupled loop, detecting (1360) one or more virtual time nodes created on the loop using a photodetector associated with the loop, and outputting (1370) a response based on the detected virtual time nodes.

[0130] In some embodiments, the reservoir receives an input signal, and injecting at least one mask input signal comprises: injecting the mask input signal into the first loop of the side-coupled loop via a directional coupler associated with the first loop.

[0131] In some embodiments, delaying at least one injected input signal includes: delaying the injected input signal by propagation on a first delay line; and coupling the delayed input signal to a next loop via a tunable directional coupler (TDC), the next loop being adjacent to the loop associated with the first delay line. Delaying at least one injected input signal also includes: delaying the coupled input signal on a second delay line associated with the next loop.

[0132] It should be understood that although specific embodiments of the technology have been described herein for illustrative purposes, various modifications may be made without departing from the scope of the technology. Therefore, the specification and drawings are to be regarded only as illustrative of the present disclosure as defined by the appended claims, and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this disclosure.

[0133] It is obvious that the above embodiments of this disclosure are exemplary and can be varied in many ways. Such present or future variations should not be considered as a departure from the spirit and scope of this disclosure, and all such modifications that are obvious to those skilled in the art are intended to be included within the scope of the following claims.

Claims

1. A multi-channel coherently driven storage tank, characterized in that, The reserve pool includes: A common mask function generator is used to generate at least one mask input signal using at least one input signal, the at least one input signal being periodically kept constant within the period of a common time mask function; Multiple side-coupled loops physically separated from each other, each loop including a delay line for delaying the at least one masked input signal, the delay lines being desynchronized in different ways with other delay lines associated with the side-coupled loops, for creating a more complex interconnect network architecture; Multiple directional couplers are associated with the side coupling loops, and at least one of the directional couplers is used to inject the at least one masked input signal into one of the side coupling loops. Multiple photodetectors, each associated with one of the side-coupled loops, are used to detect one or more virtual time nodes created on the loop.

2. The storage pool according to claim 1, characterized in that, The delay time associated with the delay line is determined based on the mask period, the number of virtual time nodes, and the offset associated with the delay line.

3. The storage pool according to claim 1, characterized in that, The reservoir also includes a tunable directional coupler (TDC) for coupling the side-coupled loops to each other.

4. The storage pool according to claim 2, characterized in that, The reservoir also includes a tunable directional coupler (TDC) for coupling the side-coupled loops to each other.

5. The storage pool according to claim 3, characterized in that, The TDC is also used to couple the at least one masked input signal delayed by the delay line to an adjacent loop.

6. The storage pool according to claim 4, characterized in that, The TDC is also used to couple the at least one masked input signal delayed by the delay line to an adjacent loop.

7. The storage pool according to claim 1, characterized in that, One or more of the common mask function generator, the plurality of side-coupled loops, the plurality of directional couplers, and the plurality of photodetectors are integrated in a silicon chip.

8. The storage pool according to claim 2, characterized in that, One or more of the common mask function generator, the plurality of side-coupled loops, the plurality of directional couplers, and the plurality of photodetectors are integrated in a silicon chip.

9. The storage pool according to claim 3, characterized in that, One or more of the common mask function generator, the plurality of side-coupled loops, the plurality of directional couplers, and the plurality of photodetectors are integrated in a silicon chip.

10. The storage tank according to claim 4, characterized in that, One or more of the common mask function generator, the plurality of side-coupled loops, the plurality of directional couplers, and the plurality of photodetectors are integrated in a silicon chip.

11. The storage tank according to claim 5, characterized in that, One or more of the common mask function generator, the plurality of side-coupled loops, the plurality of directional couplers, and the plurality of photodetectors are integrated in a silicon chip.

12. The storage pool according to claim 6, characterized in that, One or more of the common mask function generator, the plurality of side-coupled loops, the plurality of directional couplers, and the plurality of photodetectors are integrated in a silicon chip.

13. The storage pool according to claim 7, characterized in that, The delay line and a plurality of tunable Bessel filters associated with the side-coupled loop for adjusting the desynchronization of the side-coupled loop are integrated in the silicon chip.

14. The storage pool according to claim 8, characterized in that, The delay line and a plurality of tunable Bessel filters associated with the side-coupled loop for adjusting the desynchronization of the side-coupled loop are integrated in the silicon chip.

15. The storage tank according to claim 9, characterized in that, The delay line and a plurality of tunable Bessel filters associated with the side-coupled loop for adjusting the desynchronization of the side-coupled loop are integrated in the silicon chip.

16. The storage pool according to claim 10, characterized in that, The delay line and a plurality of tunable Bessel filters associated with the side-coupled loop for adjusting the desynchronization of the side-coupled loop are integrated in the silicon chip.

17. The storage pool according to claim 11, characterized in that, The delay line and a plurality of tunable Bessel filters associated with the side-coupled loop for adjusting the desynchronization of the side-coupled loop are integrated in the silicon chip.

18. The storage tank according to claim 12, characterized in that, The delay line and a plurality of tunable Bessel filters associated with the side-coupled loop for adjusting the desynchronization of the side-coupled loop are integrated in the silicon chip.

19. The storage tank according to claim 13, characterized in that, The delay line is a helical waveguide.

20. The storage tank according to claim 14, characterized in that, The delay line is a helical waveguide.

21. The storage tank according to claim 15, characterized in that, The delay line is a helical waveguide.

22. The storage tank according to claim 16, characterized in that, The delay line is a helical waveguide.

23. The storage pool according to claim 17, characterized in that, The delay line is a helical waveguide.

24. The storage tank according to claim 18, characterized in that, The delay line is a helical waveguide.

25. The storage tank according to claim 13, characterized in that, Each tunable Bessel filter includes a coupled ring waveguide (CROW), which is controlled using one or more integrated heaters.

26. The storage tank according to claim 14, characterized in that, Each tunable Bessel filter includes a coupled ring waveguide (CROW), which is controlled using one or more integrated heaters.

27. The storage pool according to claim 15, characterized in that, Each tunable Bessel filter includes a coupled ring waveguide (CROW), which is controlled using one or more integrated heaters.

28. The storage pool according to claim 16, characterized in that, Each tunable Bessel filter includes a coupled ring waveguide (CROW), which is controlled using one or more integrated heaters.

29. The storage pool according to claim 17, characterized in that, Each tunable Bessel filter includes a coupled ring waveguide (CROW), which is controlled using one or more integrated heaters.

30. The storage tank according to claim 18, characterized in that, Each tunable Bessel filter includes a coupled ring waveguide (CROW), which is controlled using one or more integrated heaters.

31. The storage pool according to any one of claims 7 to 30, characterized in that, The reservoir also includes a micro-ring resonator (MRR) based add-drop filter for coupling the side-coupled loops to each other.

32. A method for operating a reservoir with multi-channel coherent drive, characterized in that, The method includes: Receive at least one input signal; During the period of the common time mask function, the at least one input signal is periodically kept constant; At least one masked input signal is generated using the at least one input signal and the common time mask function; The at least one mask input signal is injected into the reservoir, the reservoir comprising a plurality of side-coupled loops, each of the side-coupled loops being physically separated from each other; At least one injected input signal is delayed using a delay line associated with each of the side-coupled loops, and each of the delay lines is synchronized with each other in a different way to create a more complex interconnect network architecture. For each of the side-coupled loops, one or more virtual time nodes created on the loop are detected using a photodetector associated with the loop; Output a response based on the detected virtual time point.

33. The method according to claim 32, characterized in that, The reservoir receives an input signal, and injecting the at least one masked input signal comprises injecting the masked input signal into the first loop of the side-coupled loop via a directional coupler associated with the first loop of the side-coupled loop.

34. The method according to claim 33, characterized in that, The delay of at least one injected input signal includes: The injected input signal is delayed by propagation on the first delay line; The delayed input signal is coupled to the next loop via a tunable directional coupler (TDC), the next loop being adjacent to the loop associated with the first delay line; The coupled input signal is delayed on a second delay line associated with the next loop.

35. The method according to claim 32, characterized in that, The reservoir simultaneously receives two or more input signals, and injecting the one or more masked input signals includes injecting each masked input signal into the corresponding loop of the side-coupled loop via a directional coupler associated with the corresponding loop.

36. The method according to claim 35, characterized in that, The delay of at least one injected input signal includes: Each injected input signal is delayed by propagating on the delay line associated with the corresponding loop; Each delayed input signal is coupled to the next loop via a tunable directional coupler (TDC), the next loop being adjacent to the corresponding loop.

37. The method according to any one of claims 32 to 36, characterized in that, The delay time associated with each delay line is determined based on the mask period, the number of virtual time nodes, and the offset associated with the delay line.

38. The method according to any one of claims 32 to 36, characterized in that, The side-coupled loops are coupled to each other using tunable directional couplers.

39. The method according to any one of claims 32 to 36, characterized in that, At least some components of the reservoir are integrated into a silicon chip.

40. The method according to claim 39, characterized in that, At least some components include the delay line and multiple tunable Bessel filters for adjusting the desynchronization of the side-coupled loop.

41. The method according to claim 40, characterized in that, The delay line is a helical waveguide.

42. The method according to claim 40, characterized in that, Each tunable Bessel filter includes a coupled ring waveguide (CROW), wherein the CROW is controlled using one or more integrated heaters.

43. The method according to claim 40, characterized in that, The side-coupled loops are coupled to each other through a micro-ring resonator (MRR) based add-drop filter.

44. A computer program comprising instructions, characterized in that, When executed by a computer, the instructions cause the computer to perform the method according to any one of claims 32 to 43.

45. A computer-readable medium including instructions, characterized in that, When executed by a computer, the instructions cause the computer to perform the method according to any one of claims 32 to 43.

Images