PHOTONIC RESERVOIR COMPUTER

Description

[0001] The present invention relates to a photonic reservoir computer.

[0002] Reservoir computing is a data processing model (framework) derived from the theory of recurrent neural networks that maps input signals into higher-dimensional computational spaces through the dynamics of a fixed nonlinear system, the so-called reservoir.

[0003] A reservoir computer comprises three basic parts: a) the input layer, which couples the input signal into a nonlinear dynamic system, b) the "reservoir," i.e., the recurrent neural network, which is static and untrained, and c) the output layer, which typically linearly combines the states of the reservoir to provide the time-dependent output signal.

[0004] The reservoir comprises nodes (also called vertices in graph theory) or neurons, as well as connections (also called edges) between the nodes, corresponding to the associated synapses. The state of each node can be understood as a function of the input variables at the input connections or the corresponding states of the neighboring (connected) nodes.

[0005] To utilize the reservoir to solve a specific task, machine learning is used to train a set of weighting factors in the output layer using a set of sample or training data such that a combination of the optical signals recorded at the output nodes approximates, as closely as possible, a target output corresponding to the sample or training data. These weighting factors determined during training are then used to generate the output signal for the actual input signals. Reservoir computers can, after training, approximate a global optimum in use without the need for a time-consuming iterative method. Reservoir computers have therefore proven very powerful for a range of complex tasks involving time-dependent data.

[0006] In summary, training the reservoir computer simply involves training a readout mechanism to capture the state of the reservoir and map this state to the desired output. The reservoir itself remains unchanged during this training approach. The reservoir is essentially an immutable distribution network. In this network, the input signal passes through the neural network over several bits (in the sense of a digital input stream), simultaneously creating an inter-bit memory within the network.

[0007] The main advantage of the reservoir computing framework is that the training of the neural network is performed only in the output layer and the reservoir dynamics are fixed. Reservoir computing is therefore particularly suitable for use in hardware applications.

[0008] Optical reservoir computing networks, or photonic reservoir computers, can achieve particularly high bandwidths and high signal processing speeds. A basic overview of photonic reservoir computing can be found, for example, in "Katumba et al., IEEE J. Sel. Top. Quantum Electron. 24, 1-10 (2018), https: / / doi.org / 10. 1063 / 1.5129762".

[0009] Previously known photonic reservoir computers with integrated waveguides are limited to planar topologies. Examples of well-known topologies are FIG. 1 shown.

[0010] As further prior art, reference is made to WO 2018 / 219838 A1, which discloses a photonic reservoir computer comprising one or more input nodes, a plurality of reservoir nodes, and one or more output nodes, wherein the input, reservoir and output nodes are photonically meshed by means of waveguides for signal transmission.

[0011] One of the objects to be achieved by the invention is to provide an improved photonic reservoir computer which has greater flexibility in use and / or can perform more complex functions.

[0012] To solve the above-mentioned problems, a photonic reservoir computer having the features of claim 1 is proposed.

[0013] The proposed photonic reservoir computer comprises an input layer containing one or more input nodes, a reservoir layer containing multiple reservoir nodes, and an output layer containing one or more output nodes. Input, reservoir, and output layers are not necessarily to be understood as actual layers in a layered structure; rather, in most embodiments, they merely represent a functional classification and assignment of the individual nodes.

[0014] The input nodes, reservoir nodes, and output nodes are optically coupled via photon waveguides, forming a meshed network. The waveguides can be integrated into an (optical) chip or comprise, for example, optical fibers.

[0015] Nodes are referred to as "coupled" if the output of one node is connected to the input of another node by means of a waveguide or multiple waveguides, either directly or indirectly (i.e., via one or more intermediate nodes) for information transmission. Nodes are referred to as "topologically adjacent," "directly connected," or, for short, simply "connected" if the output of one node is connected to the input of another node by means of a waveguide, directly transmitting information, i.e., without an intermediate node.

[0016] Reservoir nodes are nodes of the reservoir network, each with at least one input and one output, where they are connected to other nodes via waveguides. Due to nonlinear effects and interference, the optical signal changes as it propagates from reservoir node to reservoir node through the reservoir network.

[0017] Input nodes are defined by having at least one input at which an optical input signal is coupled into the reservoir network.

[0018] Output nodes are those nodes at which the signal is read out, weighted and made available for further processing.

[0019] According to the invention, at least some of the waveguides span or cross other waveguides and / or the reservoir nodes in a non-planar manner. The spanning waveguides are spatially above or below the spanned

[0020] Waveguides or reservoir nodes are arranged. Intersecting waveguides, in turn, have a common section that serves to transport light or information between both waveguides. In graph theory terms, the topology of the network is therefore non-planar.

[0021] In other words, the invention provides for reservoir nodes to be directly connected to other reservoir nodes that, in a planar topology, do not belong to the spatially directly surrounding / neighboring reservoir nodes. The spanning / crossing waveguides thus form a three-dimensional or higher-dimensional network or a higher-dimensional network topology, even if the reservoir nodes are arranged in a two-dimensional plane, for example, for manufacturing reasons.

[0022] Due to the high propagation speed of light, the difference in the length of the connection through the spanning / crossing waveguides can usually be neglected compared to the connections to directly adjacent nodes. It is only of minor importance. If necessary, however, longer waveguide connections can optionally be compensated for by correspondingly shorter delay lines dimensioned to achieve comparable propagation times to those to directly adjacent nodes.

[0023] The more nodes an input signal reaches in an optical reservoir computing network, the more complex the interference pattern resulting from the input signal can become at the outputs.

[0024] At the same time, the input signal at each reservoir node is divided among the node's respective outputs. If the reservoir node is designed as a passive optical component, meaning it does not have an amplifier that amplifies the signal and outputs it at the outputs, the intensity decreases significantly with each additional node the signal passes through in the network. Therefore, after passing through just a few nodes, the signal has already decayed to such an extent that the interference pattern is no longer significantly altered by the subsequent reservoir nodes. This limits the complexity of the tasks that can be solved using photonic reservoir computers without on-chip amplifiers.

[0025] The invention enables a more homogeneous distribution of the input signal across the entire network and, as a result, a greater complexity in the tasks to be solved by the reservoir computer due to the resulting more complex interference patterns.

[0026] It should be noted that the features listed individually in the claims can be combined with one another in any technically reasonable manner and demonstrate further embodiments of the invention. The description further characterizes and specifies the invention, particularly in conjunction with the figures.

[0027] It should also be noted that a conjunction "and / or" used herein between two features and linking them together is always to be interpreted in such a way that in a first embodiment only the first feature can be present, in a second embodiment only the second feature can be present and in a third embodiment both the first and second features can be present.

[0028] Further advantageous embodiments of photonic reservoir computers according to the invention result from the features specified in the subclaims and those described below.

[0029] Preferably, a plurality of the reservoir nodes can each have at least one first and one second topological neighboring node connected by waveguides. The first topological neighboring node can be spatially arranged directly adjacent to the respective reservoir node, while the second topological neighboring node is spatially spaced from the associated reservoir node such that at least one intermediate node lies between the respective reservoir node and the second topological neighboring node. The intermediate node is preferably not a topological neighboring node of the respective reservoir node. This arrangement allows a more homogeneous distribution of the input signal across the reservoir layer and thus more complex interference patterns at the output node.

[0030] In several embodiments, the reservoir nodes each have between two and seven inputs and outputs. Preferably, the reservoir nodes each have between three and five inputs and outputs.

[0031] The reservoir nodes can each have between one and five inlets and between one and five outlets. Preferably, the reservoir nodes each comprise between one and four inlets and between one and four outlets, preferably between one and three inlets and between one and three outlets. Particularly preferably, the reservoir nodes can have two or three outlets.

[0032] In other words, the degree of the nodes, i.e., the number of associated adjacent edges, can, for example, preferably be between two and seven, between three and five, and especially four. The network operates particularly advantageously, for example, efficiently, when all nodes have the same degree.

[0033] A larger number of inputs and especially outputs leads to a significant loss of intensity of the optical signal as it propagates through the network. While the number of inputs and outputs generally increases the number of connections between nodes, and consequently, the average number of nodes through which the optical signal must be transported from the input nodes to the output nodes decreases, the intensity of the optical signal decreases so drastically with each node that after just a few nodes, the signal intensity falls below the noise floor, making meaningful signal analysis impossible.

[0034] In advantageous embodiments, the majority of reservoir nodes each have two inputs and two outputs. One of the inputs and / or one of the outputs of a reservoir node can be directly connected to spatially adjacent reservoir nodes via waveguides, while preferably the second input and / or the second output of the reservoir node is connected to spatially more distant reservoir nodes. In this way, the optical (input) signal is distributed more homogeneously or more widely within the network of reservoir nodes.

[0035] The reservoir nodes can be coupled to each other in pairs, either directly or via a minimal number of intermediate reservoir nodes, with the minimal number of intermediate reservoir nodes preferably corresponding to a maximum of the square root of the total number of reservoir nodes present. In other words, the reservoir layer has a topology with a comparatively small diameter.

[0036] In some designs, the reservoir computer features integrated optics. The integrated optics encompass at least the reservoir nodes. However, the integrated optics can also include the input and / or output nodes, as well as the waveguides. The integrated optics can form a photonic integrated circuit, a so-called "photonic integrated circuit" (PIC). As an integrated optics, the reservoir computer can be manufactured particularly small and very economically in large quantities.

[0037] According to a further advantageous aspect, the reservoir computer can comprise an optical interposer. The optical interposer can have waveguides for coupling the input, reservoir, and / or output nodes. The optical interposer can be a glass interposer. An optical interposer can be used to connect a photonic integrated circuit to other optical and / or electrical components. An optical interposer can be used to communicatively couple and mechanically connect multiple photonic integrated circuits.

[0038] Optionally, the majority of reservoir nodes may not include optical amplifiers. Preferably, the majority of reservoir nodes may include only passive optical elements.

[0039] For specific tasks, the reservoir layer can be implemented with a completely passive linear network, i.e., a network without amplification or nonlinear elements. The required nonlinearity is then introduced at the readout point, typically with a photodetector.

[0040] In addition to the reduced manufacturing complexity and the resulting cost advantages, another advantage of such a passive architecture is its very low power consumption, since the computation itself requires no external power. According to a further advantageous aspect, the reservoir computer can have a plurality of input nodes. The input nodes can be distributed among the reservoir nodes in the network. Additionally, the reservoir computer can have a beam splitter. The beam splitter can be configured to split an optical input signal among the input nodes.

[0041] In these drawings show schematically: FIG. 1 shows a known topology for optical reservoir computing networks; FIG. 2 shows another known topology for optical reservoir computing networks; FIG. 3 shows a topology for a reservoir computer according to the invention; FIG. 4 shows another topology for a reservoir computer according to the invention; FIG. 5 shows an alternative topology for a reservoir computer according to the invention; FIG. 6 shows another alternative topology for a reservoir computer according to the invention; FIG. 7 shows yet another alternative topology for a reservoir computer according to the invention; and FIG. 8 shows a photonic reservoir computer according to an embodiment of the invention.

[0042] The FIG. 1 and FIG. 2 show known topologies for photonic reservoir computers from the state of the art. FIG. 1 The topology shown is also known as the Swirl architecture. FIG. 2 The topology shown is called a four-port architecture.

[0043] The arrowheads of the connections between the nodes indicate the general direction of information flow. However, it should be noted that waveguides typically operate bidirectionally, meaning that, depending on the configuration of the reservoir nodes and the choice of input nodes, information flows can reverse or even occur in both directions simultaneously.

[0044] In both cases, the swirl architecture and the four-port architecture, two-dimensional mesh topologies are used, in which the reservoir nodes are always connected only to the nearest or surrounding reservoir nodes. This type of topology is also called a planar topology.

[0045] The swirl architecture has the disadvantage of containing nodes that are not symmetric. Therefore, these nodes suffer from modal radiation at each 2×1 combiner. Radiation losses are unavoidable in these single-mode combiners, and averaged across the various input phases, 50% of the power is radiated there.

[0046] Furthermore, focusing on the mixing behavior of the Swirl architecture, there are some nodes that do not contribute significantly to the dynamics. For example, nodes 00, 03, 12, and 15 in the corners play only a very limited role in distributing the input signal in the reservoir, as they are essentially just delay lines that simply distribute the power to another node.

[0047] Likewise, the other nodes 01, 02, 04, 07, 08, 11, 13 and 14 at the edge of the reservoir, which have a smaller number of neighboring nodes, contribute less to the distribution of the input signal than the nodes 05, 06, 09, 10 in the middle.

[0048] To solve the two problems of losses and mixing, the so-called four-port architecture was developed from FIG. 2 This architecture avoids the lossy 2×1 combiners and instead uses only 2×2 devices, and additional waveguides are added to increase mixing. Simulations of this four-port architecture for a 64-node network show that losses can be an order of magnitude lower than with the Swirl architecture. These simulations also show that power is more evenly distributed across the reservoir than with the Swirl architecture.

[0049] In FIG. 3 A non-planar topology for a reservoir computer 91 according to the invention is shown. It comprises one or more input nodes 04, 13, 15, several reservoir nodes 00-24, and one or more output nodes 00, 02, 07, 10, 14.

[0050] The output nodes are in FIG. 3 Each node is indicated by arrows pointing away from the node. It should be noted, however, that the position and / or weighting of the output nodes can be selected during training on a task-specific basis. In principle, all reservoir nodes 00-24 can be considered as output nodes, as long as they have a corresponding output.

[0051] The input nodes 04, 15, 17 are in FIG. 3 can be recognized by the fact that they are fed with the input signal from a schematically illustrated optional beam splitter 93. The input nodes can also essentially be any reservoir nodes 00-24 that have a corresponding free input for the input signal. The input nodes 04, 15, 17 are arranged distributed among the reservoir nodes in order to feed the input signal into the reservoir network at various points in the reservoir layer or reservoir. For reasons of clarity, input and output nodes are no longer separately identified in the following figures. However, the teaching described here can easily be transferred to the other topologies shown.

[0052] The input, reservoir, and output nodes 00-24 are photonically meshed for signal transmission by waveguides 92, 92b, 92c. According to the invention, at least one of the waveguides 92b spans at least one other of the waveguides 92c and / or at least one of the reservoir nodes 00-24.

[0053] As in the following figures FIG. 4 bis FIG. 8 A majority of reservoir nodes 00-24 in FIG. 3 each have at least one first and one second topological neighboring node connected by waveguides. By way of example, the topology shall be illustrated using an arbitrary reservoir node 07 in FIG. 3 be described in more detail.

[0054] The first topological neighboring node 02 is spatially arranged directly adjacent to the respective reservoir node 07. The second topological neighboring node 17 is spatially spaced from the respective reservoir node such that at least one intermediate node 12 lies between the respective reservoir node 07 and the second topological neighboring node 17. The intermediate node 12 is not a topological neighboring node of the respective reservoir node 07.

[0055] Generally, there are enough nodes in the reservoir computer 91 to realize the aforementioned spanning or crossing connections.

[0056] The reservoir nodes 00-24 each have between two and seven inputs and outputs in total, in particular the reservoir nodes 00-24 each have between three and five inputs and outputs in total.

[0057] Reservoir nodes 00-24 each comprise between one and five inputs and between one and five outputs, or between one and four inputs and between one and four outputs. Reservoir nodes 00-24 preferably have between one and three inputs and between one and three outputs. Specifically, the reservoir nodes comprise two or three outputs.

[0058] A plurality of reservoir nodes 00-24 each comprise two inputs and two outputs. One of the inputs and / or one of the outputs of a reservoir node is connected to spatially adjacent reservoir nodes via waveguides 92. The second input and / or the second output of the reservoir node is in turn connected to spatially more distant reservoir nodes via waveguides.

[0059] Reservoir nodes 00-24 are connected to each other in pairs, either directly or via a minimal number of intermediate reservoir nodes. The minimal number of intermediate reservoir nodes preferably corresponds to a maximum of the square root of the total number of reservoir nodes.

[0060] The topologies of the figures FIG. 3 bis FIG. 7 The only difference is which reservoir nodes are specifically connected. Depending on the application, one or the other topology may be preferable.

[0061] FIG. 8 shows a reservoir computer 91 in a highly simplified sectional view. The reservoir computer has a topology according to the invention, as shown in one of the figures FIG. 3 bis FIG. 7 shown on.

[0062] The reservoir computer 91 comprises at least one integrated optics 81, 82. The integrated optics includes at least the reservoir nodes 00-24. The integrated optics 81, 82 form a photonic integrated circuit.

[0063] Furthermore, the reservoir computer 91 includes an (optional) optical interposer 83. The optical interposer includes waveguides 88-90, 92 for coupling the input, reservoir and / or output nodes 00-24.

[0064] This allows waveguides 84-87, 88-90 to be guided in the integrated optics 82, 83 as well as in the optical interposer 83, so that waveguides 84 / 88, 85 / 88, 86 / 90, 87 / 90, 92b / 92c / see also examples FIG. 3 ) can cross / span without conflict despite the planar arrangement of the reservoir nodes 00-24 and thus form a network with a non-planar topology.

[0065] Optionally, the majority of reservoir nodes 00-24 do not include optical amplifiers. Thus, the majority of reservoir nodes 00-24, or advantageously even all reservoir nodes, have only passive optical elements.

[0066] The project was funded by the Federal Ministry of Education and Research under grant number 16KIS1291.

Claims

1. A photonic reservoir computer (91) comprising one or more input nodes (04, 13, 15), multiple reservoir nodes (00-24) and one or more output nodes (00, 02, 07, 10, 14), wherein the input nodes, reservoir nodes and output nodes (00-24) are photonic and coupled for signal transmission by means of waveguides (92, 92b, 92c), characterized in that at least one of the waveguides (92b) spans or crosses at least one other waveguide (92c) and / or at least one of the reservoir nodes (00-24) in a non-planar manner.

2. The photonic reservoir computer (91) according to claim 1, wherein a plurality of reservoir nodes (00-24) has at least a first and a second topological neighbouring node connected by means of waveguides, wherein the first topological neighbouring node is arranged spatially immediately adjacent to the respective reservoir node, and wherein the second topological neighbouring node is spatially arranged spaced apart from the associated reservoir node, such that at least one intermediate node lies between the respective reservoir node and the second topological neighbouring node, in particular wherein the intermediate node is not a topological neighbouring node of the respective reservoir node.

3. The photonic reservoir computer (91) according to any one of the preceding claims, wherein the reservoir nodes (00-24) each have cumulatively between two and eight inputs and outputs, in particular wherein the reservoir nodes (00-24) each have cumulatively between three and five inputs and outputs.

4. The photonic reservoir computer (91) according to any one of the preceding claims, wherein the reservoir nodes (00-24) each have between one and five inputs and between one and five outputs, in particular wherein the reservoir nodes each have between one and four inputs and between one and four outputs, more preferably wherein the reservoir nodes have between one and three inputs and between one and three outputs, particularly preferably wherein the reservoir nodes have two or three outputs.

5. The photonic reservoir computer (91) according to any one of the preceding claims, wherein a plurality of reservoir nodes (00-24) each has two inputs and two outputs, in particular wherein one of the inputs and / or one of the outputs in each case of a reservoir node is connected to spatially neighbouring reservoir nodes by means of waveguides (92), preferably wherein the second input and / or the second output of the reservoir node is connected to spatially more distant reservoir nodes by means of waveguides.

6. The photonic reservoir computer (91) according to any one of the preceding claims, wherein the reservoir nodes (00-24) are paired and directly coupled through a minimal number of intermediate reservoir nodes to one another, wherein the minimal number of intermediate reservoir nodes corresponds at most to the square root of the total number of reservoir nodes present.

7. The photonic reservoir computer (91) according to any one of the preceding claims, wherein the reservoir has integrated optics (81, 82) and the integrated optics at least comprise the reservoir nodes (00-24), in particular wherein the integrated optics form a photonic integrated circuit.

8. The photonic reservoir computer (91) according to any one of the preceding claims, wherein the reservoir computer comprises an optical interposer (83), wherein the optical interposer has waveguides (88-90, 92) for coupling the input nodes, reservoir nodes and / or output nodes (00-24).

9. The photonic reservoir computer (91) according to any one of the preceding claims, wherein a plurality of reservoir nodes (00-24) has no optical amplifier, in particular wherein the plurality of reservoir nodes has only passive optical elements.

10. The photonic reservoir computer (91) according to any one of the preceding claims, comprising a plurality of input nodes (04, 13, 15), wherein the input nodes are distributed between the reservoir nodes, in particular wherein the reservoir computer includes a beam splitter (93) and the beam splitter is designed to split an optical input signal to the input nodes (04, 13, 15).