Circuit arrangement for processing signals with a neural network

Description

[0001] The invention relates to a circuit arrangement for processing signals with a neural network having a plurality of nodes, wherein a number of nodes form a common logic layer and each node of a group of nodes of a first logic layer is connected to each node of a group of nodes of an adjacent second logic layer, and wherein the nodes of the first logic layer have transmitting elements configured for emitting electromagnetic radiation and the nodes of the second logic layer have receiving elements configured for receiving electromagnetic radiation, wherein the circuit arrangement has the further features of claim 1.

[0002] Neural networks for signal processing are well-known. However, implementing neural networks at the hardware level poses a challenge, particularly when a large number of connections are required. The complexity of the network increases quadratically with the number of adjacent nodes (neurons), making the scalability of such networks problematic.

[0003] Training a hardware-implemented neural network also becomes difficult with an increasing number of parallel weighted connections in conventional electronic circuits due to the physical properties of conductor tracks on an integrated circuit (chip).

[0004] DE 10 2019 220 145 A1 describes a method for configuring a neural network trained on inference hardware to compensate for hardware / software errors in the inference hardware. This involves determining a deviation between the output data of a training hardware and the output data of the inference hardware, as well as noise parameters, to ensure bit-identical results on both the training and inference hardware.

[0005] US 6,513,023 B1 describes a circuit for a neural network with a multitude of charge storage devices representing the weight of an associated network node. This allows weights to be stored by storing charge on a multitude of capacitances of switching elements, and a multitude of node output signals to be generated in accordance with a transfer function and the stored weights.

[0006] US 2022 / 0044092 A1 discloses an implementation of layers of a neural network with optical linear neurons (nodes) implemented with an electronic circuit as an interface between the layers of the neural network. Optical input signals are fed into individual interferometer branches using wavelength multiplexing. Each branch has an input cell and a weighting cell with a demultiplexer that splits the multiplexed signals into their wavelength components. A multiplexer then recombines these wavelength components into a multiplexed signal at the output of either the input cell or the weighting cell. An optical amplitude modulator overlays the absolute value of the respective node weight onto the optical input signal.

[0007] EP 3 910 690 A1 describes a neuromorphic device with a semiconductor nanostructure in which a double-barrier quantum well region and a light-emitting region are arranged between an emitter region and a collector region of the semiconductor nanostructure. Charge carriers can tunnel through the double-barrier quantum well region, and the light-emitting region can emit light in response to the flow of charge carriers.

[0008] US 2021 / 0192330 A1 discloses a neuron with a light-emitting diode, wherein at least one optical connection exists between the optical output of the neuron and an optical input of a second neuron. The optical input signals detected by photosensitive components are electrically weighted and summed in a node. For this purpose, a neuron has a plurality of photosensitive components, each of which is coupled by an optical waveguide to a corresponding light-emitting diode of a preceding node.

[0009] US 11,144,821 B2 discloses an implementation of a neuromorphic computing system with optical neurons, which has a photodetector for converting an optical excitation signal into a corresponding electrical input signal and light-emitting transmitting elements for converting an electrical output signal into a corresponding optical output signal.

[0010] US 2020 / 0250534 A1 describes an optoelectronic computing system with an optical phase to the waveguide. The optical signal is distributed to multiple nodes. Multiplexing across multiple wavelengths occurs using a single optical fiber, with only one node ever connected to multiple receiving nodes. Numerous parallel waveguides are present.

[0011] US 11,238,336 B2 discloses a circuit for an optical neural network in which a common optical waveguide is used to transmit a WDM signal with different wavelengths, the "wavelength-division multiplexer" (WDM) being a mapping unit.

[0012] US 2019 / 0065941 A1 discloses a neuro-processing computing system in which the individual layers are coupled to each other via vertical optical paths. Individual direct one-to-one connections are always provided through a set of vertical optical vias. Several layers are combined into a 3D stack, with communication occurring via individual optical vias.

[0013] Based on this, the object of the present invention is to create an improved circuit arrangement for processing signals with a neural network, which can in particular be used as a neuromorphic system and enables a significant number of interconnected nodes of great complexity.

[0014] The problem is solved by the circuit arrangement with the features of claim 1. Advantageous embodiments are described in the dependent claims. It is proposed that the group of transmitting elements of the first logic layer be coupled to the associated group of receiving elements of the adjacent second logic layer via a common passband that transmits electromagnetic radiation, in order to transmit radiation emitted by each transmitting element to all coupled receiving elements.

[0015] A group of transmitting and / or receiving elements of nodes of a logical layer is arranged in a two-dimensional matrix on a substrate, wherein each substrate forms a logical layer of the neural network, and substrates stacked on top of each other form the neural network with the interconnected logical layers. The common transmission area is located between each pair of stacked substrates and is formed by a cavity containing air or by an optical element made of light-conducting material. According to the invention, control units for weighting the emitted radiation and first and second assignment units for assigning the emitted signals to the transmitters are also provided.

[0016] By using a shared passband that transmits electromagnetic radiation, all transmitting elements are connected to all receiving elements coupled across that passband, eliminating the need for individual optical or electrical connections. This enables highly complex networking in a very small space. Furthermore, using a shared passband that receives and transmits all the electromagnetic radiation emitted by the nodes for signal transmission has the advantage of allowing it to act as a summing element. The radiation coupled in by the transmitting elements (i.e., the excitation signals) is aggregated in the shared passband and transmitted as a summed signal to each connected receiving element.

[0017] In neural networks, however, the excitation signals must be weighted individually. Due to the shared passband, the necessary discrimination or sender-receiver assignment is eliminated, since the summation occurs before the receiving elements.

[0018] Therefore, the transmitting elements are connected to a control unit that defines a characteristic of the emitted radiation by assigning a learned weight to each transmitting element (i.e., each node). The weighting of the individual emitted signals is thus determined by a characteristic of the radiation. This can be at least one characteristic, selected individually or in combination from the group of: luminous intensity (i.e., radiant power), wavelength, polarization, pulse characteristics such as pulse length, frequency, phase, and the like.

[0019] Furthermore, a first assignment unit is connected to the control units of the transmitting elements, and a second assignment unit is connected to the receiving elements. These assignment units are configured to assign the signals emitted by the transmitting elements to the receiving elements, which are coupled via the common passband. The transmit-receive assignment is thus performed by assignment units of the transmitting and receiving units, i.e., in the first logic layer and the adjacent second logic layer coupled to it. This enables the assignment of the electromagnetically radiated excitation signals of the individual transmitting units to their respective transmitting nodes at each receiving node, despite the upstream aggregation of the signals or their transmission via a common passband.

[0020] For example, modulation of the excitation signals can be performed in the transmitting logic layer and demodulation in the receiving logic layer. The first mapping unit can be configured to modulate the excitation signals emitted by the transmitting elements, and the second mapping unit can be configured to demodulate the corresponding modulated signals received by the receiving elements.

[0021] The modulation can be, for example, frequency modulation, phase modulation, amplitude modulation, wavelength modulation or pulse modulation (e.g. pulse width modulation or pulse code modulation).

[0022] However, an allocation via multiplex control is also conceivable. For this purpose, the first allocation unit can be configured for multiplexing the transmitting elements, and the second allocation unit can be configured for demultiplexing the signals acquired by the receiving elements, corresponding to the multiplexing control.

[0023] For example, the individual nodes of a sending logic layer can be addressed sequentially to transmit their respective weighted excitation signals. The receiving elements of the nodes in the receiving logic layer can then receive the excitation signals from the individual sending nodes in succession and assign them accordingly. This multiplexing process can then be performed sequentially for each node until all nodes of a logic layer have transmitted their excitation signals to the subsequent receiving layer.

[0024] In this multiplex operation, which is carried out with respect to the transmitting nodes, the summing property of the common passband is eliminated. The summation of the weighted excitation signals for the individual receiving nodes can then be performed at the electrical connection of the light receiving elements, for example by charging a capacitor.

[0025] When multiplexing is performed with respect to the receiving node, the signals from all transmitting elements of the coupled nodes of the transmitting logic layer are simultaneously transmitted across the common passband, each with a weight relative to the receiving node, and summed. This preserves the link-specific weights between the current receiving node and the individual transmitting nodes, and these weights are incorporated into the summed received signal through the summation of the weighted excitation signals. This process can then be repeated for each receiving node.

[0026] The mapping units, with their connected electromagnetic radiation-emitting (i.e., radiating) transmitters, can be configured to map the wavelength-dependent signals emitted by each transmitter to the nodes that have receivers coupled via the common passband. The transmitters can be, for example, light-emitting diodes or laser diodes, in particular vertically emitting laser diodes (VICEL - Vertical Cavity Surface Emitting Laser).

[0027] This allows the wavelength of the individual transmitting units to be utilized and regulated in such a way that an assignment between the emitted optical excitation signal and the respective optical receiver in the receiving node is possible, at least in order to be able to assign to each receiving node the summed signal of the coupled transmitting elements of the transmitting nodes intended for that node.

[0028] Using such a wavelength-dependent assignment, it is possible, for example, to transmit the excitation and de-excitation of nodes, if a certain wavelength represents an excitation signal and another wavelength represents a de-excitation signal.

[0029] Narrowband LEDs are suitable for this purpose, as they each emit a specific wavelength with a narrow bandwidth. Vertically emitting laser diodes (VICEL), whose wavelength can be controlled depending on the current, are also suitable. In this case, the wavelength is primarily dependent on the temperature, which in turn is influenced by the control current.

[0030] The control unit can have a control transistor connected to a corresponding light-emitting transmitter. Depending on the control voltage or current applied to the control transistor and the transistor's characteristic curve, the control unit can then cause the emission of a predetermined light power. The transistor's characteristic curve is thus used to determine the weight of the emitted excitation signal. The transmitter is driven by the control transistor with the gain specified by the characteristic curve, so that the radiant intensity of the excitation signal can be controlled by the transistor's characteristic curve. The control voltage or current then represents the learned weight, and this control voltage or current is related to the transistor's characteristic curve, which influences the gain factor.

[0031] The control unit can have a memory to store the control signals used as learned weights for the respective node to control the associated transmitters. For example, the control voltages or control currents used to control the associated control transistors of a node can be stored for this purpose.

[0032] Such a memory can, for example, be designed as a storage capacitor in which the learned weights are stored as a control voltage. This storage capacitor can be connected, for instance, to a corresponding control transistor of a transmitting node to control a transmitter element. The storage capacitor can include an additional charge conservation circuit, which replenishes the stored charge if it decreases due to parasitic discharge or by reading the storage capacitor.

[0033] However, it is also conceivable to use DRAM memory to store the learned weights. The weights can also be stored in a software program of a microcontroller or microprocessor, which can optionally control them with a computer program, possibly using a digital-to-analog converter to control, for example, control transistors.

[0034] The control unit can be configured for pulse-width modulation (PWM) of the transmitters connected to it. Such pulse-width modulation allows the radiated power of the individual excitation signals to be precisely controlled. Furthermore, pulse-code modulation can also be performed to enable transmit-receive assignment.

[0035] The neural network can have first-layer logical nodes, each with one transmitting element and one receiving element, for bidirectional signal transmission. These nodes are connected to the same transmitting passband. Furthermore, second-layer logical nodes, each with one receiving element and one transmitting element, are also present. This allows the nodes of the first-layer logical network to be used not only as transmitting nodes, as previously described, but also simultaneously as receiving nodes. Similarly, the second-layer logical network nodes can be used not only as receiving nodes but also as transmitting nodes.

[0036] Thus, the previously described first logical layer also has the functionality of the previously described second logical layer, and vice versa. The designation as first and second layer should therefore not be understood as referring to transmit or receive layers, but merely as a distinction between two interconnected (chained) layers, each with a number of nodes as the logical layer, or, in implementation with stacked (i.e., layered) substrates as the physical layer.

[0037] At least one layer can have a pair of transmitting elements and receiving elements, which are coupled in opposite directions to a common passband and are interconnected to at least one receiving element of an adjacent layer for the weighted forwarding of the excitation signals received by the receiving element via the transmitting element and the passband coupled to it.

[0038] This means that nodes of at least one layer are designed to receive radiated excitation signals at the input and emit excitation signals at the output as electromagnetic radiation. They are thus coupled at their input and output with a passband each.

[0039] The nodes can also be bidirectional, so that there is a pair of sending element and receiving element on each of the opposite sides.

[0040] The neural network can have at least one layer which, in addition to the group of nodes coupled to a common passband to a neighboring layer, has at least one further node which is connected to the neighboring layer via a further passband for forwarding the excitation signals in the form of emitted electromagnetic radiation.

[0041] When the claims and description refer to a number of nodes in a layer, this does not necessarily mean all nodes (neurons) in the layer. Rather, it refers only to the group of nodes in the layer that are of interest. Additional purely electrically functioning nodes or additional groups (stages) of optically functioning nodes are therefore not excluded and can be advantageously included as an extension of the neural network.

[0042] This allows, for example, different properties of the neural network to be processed and learned separately, at least in individual stages, without these properties being mixed with other properties in the areas of the neural network.

[0043] In this context, "optical" refers to any electromagnetic radiation, as distinct from conducted electrode transport. Such "optical" electromagnetic radiation includes the visible (400 to 800 nm) and invisible wavelength ranges, especially ultraviolet (100 to 400 nm) and infrared (800 nm to 1 mm), as well as X-rays (especially 1 to 100 nm). This encompasses not only optical radiation in the wavelength range of 100 nm to 1 mm according to DIN 5031. Higher-frequency radiation, especially radio waves (above 1 mm), is less suitable due to its greater propagation distance and the need for shielding. Transition to terahertz radiation (30 µm to 3 mm) is conceivable, but currently very complex.

[0044] The transmission zone is designed as a cavity for transmitting electromagnetic radiation from the group of transmitting elements in the first layer to the group of receiving elements in the adjacent second layer, or as a light-guiding optical element, such as a film or plate transparent to the wavelength of the emitted radiation. However, the transmission zone is not a one-to-one connection between a transmitting element and an associated receiving element, but rather a shared area that connects a group of transmitting elements in a first layer with a group of receiving elements in an adjacent second layer. A cavity, film, or plate can achieve a homogeneous distribution of the electromagnetic radiation.For electromagnetic radiation in the visible and invisible optical wavelength range and the X-ray range, point emitters can be realized with a beam angle and aspect ratio that allow a small distance between the adjacent layers, i.e., a small thickness of the transmission area in the millimeter range (<1 mm to about 5 mm).

[0045] The passband can optionally be modified in sections to influence the radiation pattern differently in different areas, for example by polarization. This allows for an unequal distribution of the emitted radiation to different subgroups of receiving elements, in order to impose matching factors on the summing function inherent in the passband for the excitation signals.

[0046] The nodes of an input layer can have electrical connections for electrical input signals, which serve to control the light-emitting transmitting elements connected to them in a weighted manner.

[0047] Optionally, or in combination with this, the nodes of an output layer can each have electrically connected terminals for electrical output signals with an associated receiving element.

[0048] This means the neural network can be configured to process electrical input signals. It can also be configured to output electrical signals as the output of the neural network.

[0049] When multiple neural networks are interconnected, an output layer with an electrical signal output can be electrically connected to the input layer of a subsequent neural network.

[0050] However, it is also conceivable to process "optical" input signals in the form of electromagnetic radiation at the input of the neural network and the

[0051] Providing "optical" output signals as a processing result of the neural network at the neural network's output. The inputs of the nodes in the input layer then have receiving elements configured to receive electromagnetic radiation, i.e., not to receive conducted electrical signals. For optional output of optical signals at the neural network's output, the nodes in the output layer have corresponding transmitting elements configured to emit electromagnetic radiation. Prior weighting may be implemented in the output layer when controlling the transmitting elements or the electrical elements. In an example not pertaining to the claimed invention, the summed excitation signals arriving at the nodes of the output layer can also optionally be passed on unweighted at the output of the respective node in the output layer.

[0052] The invention is explained in more detail below with reference to exemplary embodiments and the accompanying drawings. These show: Figure 1 - Sketch of a three-layer neural network; Figure 2 - Schematic diagram of the weighting and summation of input signals in a node with an additional function for generating an output signal; Figure 3 - Circuit diagram of a transmitter element in the form of a light-emitting diode with a weighted control signal; Figure 4 - Diagram of the product obtained from factors A and B when controlling the transistors as the weighting result; Figure 5 - Circuit diagram for a receiver of a node for receiving electromagnetic radiation; Figure 6 - Sketch of a neural network formed from matrix substrates with integrated micro-transmitter / receiver elements and an intermediate film; Figure 7 - Sketch of a three-layer neural network setup with input and output signals.

[0053] Figure 1 Figure 1 shows a sketch of a circuit arrangement for processing signals with a neural network, which consists by way of example of three logical layers S 1 , S 2 , S 3, which are formed according to the invention by substrates stacked on top of each other in a planar fashion.

[0054] Each logical layer S i has a plurality of nodes K i, 1, K i, 2, K i, 3 ​​.

[0055] Each node K i, j forms a neuron, with the connections (synapses) between the neurons in the depicted section of the neural network being made by light.

[0056] For this purpose, the first logic layer S1, for example, has node K1,1 as its input layer, containing electromagnetic radiation-emitting transmitting elements such as light-emitting diodes (LEDs). According to the invention, a micro-LED array is used in which LEDs for the transmitting elements and / or photodiodes / phototransistors for the receiving elements are integrated on a substrate together with the control electronics as a semiconductor device. The transmitting and / or receiving elements are arranged on the substrate in a two-dimensional matrix. In particular, for the realization of an intermediate layer, a matrix of transmitting and / or receiving elements can be arranged on both the top and bottom surfaces of the substrate to connect to an adjacent layer located below and above the substrate.

[0057] The output signals of the neurons of the input layer S1 are generated using such a group of transmitting elements designed to emit electromagnetic radiation.

[0058] The input of these nodes K i, 1 can either be an electrical input for receiving electrical signals or an optical input for receiving input signals in the form of electromagnetic radiation (also called excitation signals).

[0059] The hidden second layer S 2 has a plurality of nodes K i, 2 , each of which has at its input a receiving element to receive the excitation signals of the nodes K i, 1 of the preceding layer and at its output a transmitting element to radiate excitation signals to the following layer.

[0060] The subsequent layer can be another hidden layer of the type of the second layer S2 with a plurality of nodes Ki,3 of the same or different number, or, as shown, already an output layer S3. The output layer has a plurality of nodes Ki,3 with receiving elements to receive the excitation signals from the preceding layer S2. The output signal OUTi,3 is then provided either as an electrical or as an optical signal, i.e., as electromagnetic radiation (especially in the wavelength range from 10 pm to 1 mm, preferably 1 nm to 50 µm).

[0061] It can be seen that the successive logical layers S1, S2 and S2, S3 have connections between each preceding node Ki,1 and each subsequent node Ki,2. The excitation signals between the individual nodes Ki,1 and Ki,2 are each weighted individually.

[0062] According to the invention, the individual connections are now realized by a common passband, which allows the emitted electromagnetic radiation to pass through and is formed by a (e.g., air-filled) cavity or a light-conducting plate, instead of individual optical fibers. In this context, "light-conducting" refers to the ability to transport electromagnetic radiation in the aforementioned wavelength range, which also includes invisible radiation, from the group of transmitting elements to the adjacent group of receiving elements in the neighboring layer. This ensures that the individual signals are summed equally for all receiving nodes when all transmitting and receiving nodes are active simultaneously. The common passband thus performs the function of summing the input signals for the receiving nodes.

[0063] This function, however, is currently implemented in the individual receiving nodes. This is due to the Figure 2 to recognize, which shows a sketch of the function of a node K i, j with a weighting of input signals (xi · wi ) with the weights wi and a summation of the weighted input signals xi · wi.

[0064] In a receiving node, the summed input signal is then adapted with a function Φ (a) to generate an output signal Y.

[0065] It becomes clear that input signals x₀, x₁, x₂, x₃, and optionally input parameters L, are each multiplied by a weight W₀, W₁, W₂, W₃, B of the respective node. These weighted input signals xi · wi are then summed (Σ) and processed in the receiving node using a function Φ(a) predefined there. The weights wi, b, and the function Φ(a) are trained in a known manner. They can be further adapted during operation or be predefined as fixed values ​​of a trained neural network.

[0066] The output signal Y can be the output signal of an output node at the end of the neural network. However, it is also conceivable that the output signal Y of a node is passed on to the nodes of the subsequent layers. In this case, the output signal Y forms an input signal xi in each subsequent layer, so that the functionality described is repeated in the following layer.

[0067] Figure 3 Figure 1 shows an exemplary circuit diagram for controlling transmitting elements D1 of a node Ki,j with a weighted control signal w = f(A, B), which varies a characteristic property of the excitation signal emitted by transmitting element D1 that can be evaluated by the receiving elements of the subsequent nodes. This property could, for example, be the radiation intensity. Thus, the weighted output signal emitted by a node can be a radiated excitation signal with a radiation intensity that depends on the magnitude of the input signal and the weighting.

[0068] The transmitting element D1 (LED) can be controlled at the optimal reference point, for example, using an LED driver LED-CTR. The LED driver LED-CTR can, for instance, control a large number of LEDs in an LED array, allowing multiple nodes of a layer to be controlled simultaneously with their respective weighted signals.

[0069] Each node is weighted, for example, by a series connection of two transistors T1 and T2, which multiplies the two input voltages A and B. These transistors T1 and T2 are connected to the supply voltage VCC and to ground GND via a resistor R1.

[0070] A voltage-time converter VTC can be provided between the weighted output signal w = f (A, B) and the LED driver to convert a drive voltage into a pulse width PB with which the LEDs are operated.

[0071] The pulse width PB can be controlled by a pulse signal as the second input signal of the second voltage-to-time converter, for example by a microprocessor.

[0072] The weighting can be performed using a transistor characteristic curve. For this purpose, two transistors T1 and T2 can be connected in series. The weighting does not correspond to an ideal multiplication and cannot be approximated as such. The behavior is a non-linear characteristic curve (product w = f(A, B)), which is shown in the diagram below. Figure 4 This is an example.

[0073] Such a nonlinear characteristic curve can still be used effectively for an "optical" neural network, even though it does not provide ideal multiplication. This nonlinearity is automatically compensated for during the training of the neural network and therefore plays no role.

[0074] This circuit arrangement shown, with weighting using two transistors connected in series and utilizing the transistor characteristic curve, enables a circuit design that requires very few components.

[0075] The circuit arrangement can, for example, be implemented as an integrated circuit in a very small space with a significant number of nodes (neurons). This is particularly helpful for a neuromorphic system that approximates biological systems, such as the human gene structure with its extremely large number of neurons connected by synapses.

[0076] Figure 5 Figure 2 shows an exemplary circuit diagram for a receiving element D2, which converts optical signals received from preceding nodes into an electrical signal via a common passband, such as a light-guiding optical element in the form of a plastic film. The emitted excitation signals of the transmitting LEDs of the preceding nodes in the previous layer are detected, summed, and converted into a current. A photodiode in solar cell mode is used for this purpose, for example.

[0077] This receiving element D 2, which is sensitive to electromagnetic radiation, is connected to the supply voltage VCC via an anode.

[0078] On the other side, the cathode is connected to the input of transistor T3. This is a switching transistor that uses a TDM (transistor detection and management) method to enable or disable the charging of a neuron. This is controlled by the control signal L (load) via a resistor R2.

[0079] The switching transistor T 3 is connected to a diode D 2, which acts as a blocking diode to prevent unwanted discharge of the subsequent capacitor C 1.

[0080] Capacitor C1 and the anode of blocking diode D2 are connected with their cathodes to the drain terminal of transistor T4, which is driven by a signal dis at its gate via resistor R3. The source terminal of transistor T4 is connected to ground (GND) via resistor R4.

[0081] This means that capacitor C 1 is actively discharged via the discharge resistor R 4 when resetting the activation is required, such as when the network calculation is completed.

[0082] The activation of the neuron, via the summation of the received pulsed light signal from the previous layer by photodiode D2, occurs during the TDM phases (Time Division Multiplex phases), in which the charging of the neuron is enabled by the released signal L (load). The energy transmitted by photodiode D1, i.e., the current flow, depending on the absorbed light energy, is stored in capacitor C1. The voltage OUT present across the capacitor after the charging process is complete then constitutes the output signal.

[0083] The control signals L and diss are generated, for example, by digital electronics (e.g., microprocessor, microcontroller) so that exactly one neuron is loaded per time slot, the activation is stored across the entire network width, and finally, a calculation is reset using the switching transistor T4 via the reset signal diss. This forms part of the mapping unit Z, which, with the help of the control signals L and diss, enables individual control of the receiving elements and thus a transmitter-receiver mapping.

[0084] Time-division multiplexing (TDM) ensures the temporal selection of neurons (nodes) in the next network layer. At any given time, exactly one neuron in the following layer is processed. This means that a light-receiving element D2 of node Kij in the receiving layer is always activated, and the light signals from the preceding nodes, summed via the common optical element, are collected in capacitor C1. The sequential timing of the TDM activation of the individual nodes in a layer can be achieved, for example, using digital electronics, such as a suitably programmed microcontroller or microprocessor, and the resulting activation of the switching transistors T3 for each node (neuron).

[0085] The circuit arrangements shown utilize an LED array. This is in the Figure 6The diagram shows a uniform, two-dimensional array of so-called micro-LEDs. The individual micro-LEDs are in the micrometer range, for example, 5–100 µm, and can be square (rectangles, circles, etc.). The entire micro-LED array contains a large number of micro-LEDs. The P-contact of each LED is connected to the outside via a conductor to electrically control the respective LED. As shown, the LEDs can be controlled at their anodes by individual control circuits. They can be directly addressed. The LEDs share a common N-contact, which is connected to the outside via at least one contact. Therefore, the LED array can be electrically controlled on a circuit board.

[0086] The individual logic layers S1, S2, S3 are formed by substrates on which a two-dimensional matrix with pairs of transmitting elements D1 and receiving elements D2 is integrated as a semiconductor device on one or both sides. These substrates can then be mounted as modules on electronic circuit boards or...

[0087] Electronic substrates are used up or integrated together with the control electronics.

[0088] The substrates are then stacked on top of each other, with the transmittance zone O formed between each stacked substrate. The transmittance zone O is realized by frame-shaped spacers to provide an air cavity or by an intermediate film or plate transparent to the radiation wavelength range. Such a film or plate then forms a light-guiding optical element. When integrating transmitting / receiving elements in the range of 10 nm to 1 µm, the aspect ratio is maintained even with thin films at a small distance of 1 to 5 mm, preferably 1 to 2 mm, to ensure a homogeneous distribution of the emitted radiation to all receiving elements of the adjacent layer. It is also conceivable to provide different transmittance zones, for example, to...Different polarizations or other region-specific changes cause the radiated excitation signals to be propagated unequally to the receiving elements of the adjacent group. This allows different factors for subgroups of transmitting elements or receiving elements to be imprinted on the summing function inherent in the passband when multiple excitation signals are simultaneously transmitted to at least one receiving element.

[0089] Figure 7 Figure 1 shows a sketch of a circuit arrangement for processing signals with a previously described neural network.

[0090] Here, three layers S1, S2, and S3 are shown as examples. However, there could just as easily be only two layers, or, with additional hidden layers, more than three layers.

[0091] The number of nodes in each layer does not have to be the same. It can be defined individually for each layer and vary from layer to layer. This is illustrated in the transition from the second layer S2 to the third layer S3, where three nodes become two nodes.

[0092] The first layer S 1 serves as the input layer. The input signals IN j are introduced there into the respective nodes xi, 1.

[0093] These can be electrical or "optical" signals.

[0094] The nodes of the input layer S1 have electromagnetic radiation-emitting transmitting elements (LEDs), which can be, for example, micro-LEDs from an LED array. Each LED is individually controlled to generate a weighted light signal for a subsequent node of the following layer S2.

[0095] The weighted output signal is given by the equation x i , j = IN i , 1 ⋅ w i , j .

[0096] The index i refers to the respective node of the first logical layer S 1 and the index j to the target node of the second logical layer S 2 .

[0097] Thus, the signal from the topmost node x1,1 of the first logic layer S1 to the topmost node x1,1 of the second logic layer S2 is weighted with w1,1. The signal from the topmost first node x1,1 to the third node of the second logic layer S2 then has the corresponding weight w1,3.

[0098] In the respective receiving layer (such as the second hidden logic layer S2), the received, weighted signals are summed for each node xi,2. This can be described by the formula: x j , 2 = ∑ i = 1 N x j , 2 ⋅ w j , 2

[0099] The index j corresponds to the index of the respective target node of the second layer S 2 .

[0100] The nodes of the second logic layer S 2 each have a receiving element at the input, such as a photodiode or a phototransistor, and a transmitting unit at the output, such as a light-emitting diode (LED).

[0101] In the last receiving layer, such as the logical layer S 3, the excitation signals are each passed through a common passband O for a node and summed up.

[0102] Since the summing function is performed in the common passband O between the respective logical layers S 1 , S 2 and S 2 , S 3 arranged one after the other, an assignment of the excitation signals of the transmitting nodes to the respective receiving nodes must be made.

[0103] This can be achieved through time-division multiplexing (TDM) control via a clock generator CU1, which sends a clock signal to the respective control units CU2 and CU3 in the downstream logic layer S1 and S3. The clock generators CU1, CU2, and CU3 together form the allocation unit Z. The receiver circuits of the individual receiver nodes in the receive layer S2, or, when transmitting from S2 to S3, in layer S3, are then activated sequentially, ensuring that only one receiver node can receive at a time. The transmitters in the node of the upstream layer S2 are then controlled by the clock signal to select the weights wi and j intended for the respective receiver nodes when the respective receiver node is activated.

[0104] The individual connecting lines shown in the sketch are actually realized through a common passage area, for example a cavity or a light-guiding plate or film, to which all associated transmitting elements (LEDs) and subsequent receiving elements are connected in the same way.

[0105] As an alternative to the time-division multiplexing method shown for assignment using the assignment unit Z (clock generator CU1, CU2, CU3), other assignment methods can also be used.

[0106] It is conceivable that the excitation signals emitted by the transmitting elements to the subsequent layer are modulated and then demodulated accordingly by receiving units in the receiving nodes of the subsequent layers. The individual signals, each weighted by a receiving node, can be, for example, frequency-modulated, phase-modulated, amplitude-modulated, wavelength-modulated, and / or pulse-modulated.

[0107] The assignment can also be made by the wavelength dependence of the individual excitation signals. Vertically emitting laser diodes such as VICEL are suitable for this purpose, as their wavelength can be variably controlled via the temperature and thus via the drive current.

[0108] The weights w used for control are trained as is well known. They can then be stored as unchangeable parameters in the circuit arrangement, for example in a DRAM memory or in a storage capacity connected to the respective control transistor.

[0109] The weights stored in this way can also be relearned during operation of the circuit arrangement.

[0110] In this embodiment, only one direction of communication is described. The circuit arrangement can also be operated in the reverse direction if the corresponding nodes are configured for bidirectional communication. On the receiving side shown, each node would then need to contain additional transmitting elements that emit electromagnetic radiation, and on the transmitting side, each node would need to contain additional receiving elements that detect electromagnetic radiation. Alternatively, the network could be duplicated and rotated, with one node connection in the forward direction and one connection from other nodes in the return direction.

[0111] The principle of summing the excitation signals for each receiving node with a common passband and assigning the summed weighted excitation signals to an assignment unit is retained even in bidirectional communication and can be applied equally for return communication.

Claims

1. A circuit arrangement for processing signals using a neural network having a plurality of nodes, wherein a number of nodes form a common logical layer, and each node of a group of nodes of a first logical layer is connected to each node of a group of nodes in an adjacent second logical layer, and wherein the nodes of the first logical layer have transmitting elements configured to emit electromagnetic radiation, and the nodes of the second logical layer have receiving elements configured to receive electromagnetic radiation, characterized in that - the group of transmitting elements of the first logical layer is coupled to the corresponding group of receiving elements of the second logical layer via a common passage that transmits electromagnetic radiation, in order to transmit radiation emitted by the transmitting elements to all coupled receiving elements, - the transmitting elements are connected to a control unit that specifies a characteristic of the emitted radiation via a learned weight, and - a first assignment unit is connected to the control units of the transmitting elements and a second assignment unit is connected to the receiving elements, wherein the assignment units are configured to assign the signals emitted by the transmitting elements to the receiving elements coupled via the common passage for electromagnetic radiation arranged, - a group of transmitting elements and / or receiving elements of nodes of a logical layer is arranged in a two-dimensional matrix on a substrate, wherein each substrate forms a logical layer of the neural network and substrates stacked flat on top of one another form the neural network with the interconnected logical layers, and wherein - the common passage is arranged between a pair of substrates lying flat on top of one another and is formed by a cavity containing a volume of air or by an optical element made of light-conducting material.

2. A circuit arrangement according to claim 1, characterized in that the first assignment unit is configured to modulate the signals emitted by the transmitting elements, and the second assignment unit is configured to demodulate the modulated signals received by the receiving units.

3. A circuit arrangement according to claim 2, characterized in that the first assignment unit is configured for frequency modulation, phase modulation, amplitude modulation, wavelength modulation, or pulse modulation.

4. A circuit arrangement according to claim 1, characterized in that the first assignment unit is configured for multiplex control of the transmitting elements, and the second assignment unit is configured for demultiplex evaluation of the signals detected by the receiving elements corresponding to the multiplex control.

5. A circuit arrangement according to one of the preceding claims, characterized in that the assignment units and the transmitting units connected thereto are configured for wavelength-dependent assignment of the signals emitted by a respective transmitting unit at the nodes that have receiving units coupled via the common passage for electromagnetic radiation, wherein the transmitting units are light-emitting diodes or laser diodes, in particular vertical-cavity surface-emitting lasers (VCSEL - Vertical Cavity Surface Emitting Laser).

6. A circuit arrangement according to one of the preceding claims, characterized in that the control unit has a plurality of drive transistors, wherein a respective drive transistor is connected to an associated transmitting element to control the characteristics of signal emission as a function of the control voltage or control current applied to the drive transistor and the transistor characteristic curve.

7. A circuit arrangement according to claim 6, characterized in that the control unit has a memory for storing the control voltages or control currents used as learned weights for the respective node to drive the associated drive transistors.

8. A circuit arrangement according to claim 7, characterized in that the learned weights are stored as control voltages in a storage capacitor connected to the respective drive transistor.

9. A circuit arrangement according to claim 7, characterized in that the learned weights are stored in a DRAM memory.

10. A circuit arrangement according to one of the preceding claims, characterized in that the control unit is configured for pulse width modulation (PWM) of the transmitting elements connected to the control unit.

11. A circuit arrangement according to any of the preceding claims, characterized in that at least one layer has nodes, each with a pair of transmitting elements and receiving elements, which are coupled in opposite directions to a respective passage for electromagnetic radiation and are configured for the weighted forwarding of the signals received by the receiving element via the transmitting element and the passage coupled thereto with at least one receiving element of an adjacent layer.

12. A circuit arrangement according to any of the preceding claims, characterized in that the neural network for bidirectional signal transmission comprises nodes of the first layer, each having a transmitting element and a receiving element coupled to the same passage, and nodes of the second layer, each having a receiving element and a transmitting element.

13. A circuit arrangement according to any of the preceding claims, characterized in that the neural network has at least one layer which, in addition to the group of nodes coupled to an adjacent layer via a common passage for electromagnetic radiation, has at least one further node that is optically connected to the adjacent layer via a further passage for electromagnetic radiation.

14. A circuit arrangement according to any of the preceding claims, characterized in that the nodes of an input layer have electrical connections for electrical input signals, which serve to provide weighted control of the respective transmitting elements connected thereto, and / or the nodes of an output layer have connections for electrical output signals that are electrically connected to a respective receiving element.

15. A circuit arrangement according to any of the preceding claims, characterized in that the optical element forming the common passage is a film or plate made of a material transparent to the radiation emitted by the transmitting elements.

16. A circuit arrangement according to one of the preceding claims, characterized in that a radiation-transmissive film, each forming the optical element of the common passage, is arranged between a pair of substrates lying flat on top of one another.