Neuromorphic photonic processor using optical low-coherence interferometry

The neuromorphic photonic processor employing low optical coherence interferometry addresses the limitations of electronic ANNs by enabling efficient, scalable, and low-energy AI processing through coherence domain multiplexing, enhancing integration density and reducing thermal management needs.

WO2024240969A1PCT designated stage expired Publication Date: 2024-11-28BERILIO INVESTIGACIONES A I E
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
PCT/ES2023/070329
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-22
Publication Date
2024-11-28

AI Technical Summary

Technical Problem

Current electronic approaches for implementing Artificial Neural Networks (ANNs) face limitations in capacity, scalability, and energy consumption, particularly due to high operating frequencies, complex interconnection requirements, and increased thermal management needs in large data centers, which are not efficiently addressed by conventional digital systems.

Method used

A neuromorphic photonic processor using low optical coherence interferometry is designed to perform multiplication-accumulation operations, leveraging coherence domain multiplexing to enable parallelization and reduce energy consumption, with a simplified detection scheme using a single detector per layer, allowing for scalable and efficient AI algorithm execution.

Benefits of technology

The solution achieves scalable and energy-efficient AI processing by enabling parallel MAC operations at the speed of light with lower latency and heat dissipation, improving integration density and reducing the complexity and cost of photonic processors, thus overcoming the limitations of traditional electronic systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure ES2023070329_28112024_PF_FP_ABST
    Figure ES2023070329_28112024_PF_FP_ABST
Patent Text Reader

Abstract

Neuromorphic photonic processor comprising: a light source (1); a multipath optical interferometer; an optical splitter (2); a plurality of M neurons (3), where each neuron (3) has N inputs defining a plurality of MxN optical paths or interferometric arms (4); an optical path (5); an optical combiner (6); a detection stage (7), an electro-optical converter (8), an electronic circuit (9), and a non-linear activation module (10); wherein the optical paths or interferometric arms (4) in turn comprise: a first optical modulator (4.1), a second optical amplitude modulator (4.2), a phase delayer (4.3), and an optical wave guide (4.4). The processor is based on low-coherence light sources, resulting in low coherence interferometry, such that all neurons can be multiplexed in the coherence domain, which simplifies the detection system, and the output of each neuron can be recovered using demultiplexing techniques with a single detector.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Neuromorphic photonic processor using low-coherence optical interferometry

[0002] DESCRIPTION

[0003] OBJECT OF THE INVENTION

[0004] The object of the present invention, as the title of the invention establishes, is a neuromorphic photonic processor using low optical coherence interferometry.

[0005] The present invention is characterized by the special design and configuration of each and every one of the elements such that a software-configurable multipurpose photonic processor is achieved that implements one of the most critical functionalities carried out by ANNs (Artificial Neural Networks); that is, multiply-accumulate (MAC) operations. The photonic processor is expected to be jointly integrated with high-speed electronics to build the fully functional architecture capable of executing AI algorithms. The invention exploits coherence domain multiplexing (CDM), which is based on low-coherence optical interferometry, to expand computational capacity through parallelization while providing an unprecedented compromise between scalability and physical resource utilization.

[0006] Therefore, the present invention falls within the scope of photonic processors.

[0007] BACKGROUND OF THE INVENTION

[0008] Artificial Intelligence (AI) is expected to drive the Fourth Industrial Revolution in combination with other relevant cross-cutting technologies. Applications such as diagnostic imaging in healthcare settings, autonomous driving enabled by advanced driver assistance systems (ADAS), accurate weather forecasting, voice-based intelligent digital assistants, automated stock trading, and natural language processing, to name a few, are currently undergoing profound changes. These changes are resulting in unprecedented levels of performance and accuracy thanks to the deployment of AI.

[0009] However, new AI algorithms are typically powered by large amounts of datasets (big data) and supported by deep artificial neural networks (ANNs). These factors motivate the use of high-performance computing systems, typically running on dedicated hardware based on graphical processing units (GPUs) and / or tensor processing units (TPUs). Despite sustained progress over the past few decades in the capacity of digital processing systems, Moore's Law is beginning to saturate as demand continues to increase. Furthermore, the model cannot be sustainably scaled due to rising energy consumption. Under this scenario, innovative, low-power solutions are required that simultaneously break the capacity bottleneck.Recently, the cointegration of photonic engines with high-speed electronics has become a preferred solution by leveraging the enormous bandwidth and low-power potential inherent in optical technologies.

[0010] Current electronic approaches, such as CPUs, GPUs, or TPUs, which implement the most advanced ANNs, have a number of limitations inherent to the technology. These limiting factors can be partially alleviated by cointegrating photonic solutions to perform some of the operations. The most significant limitations of state-of-the-art electronic approaches are:

[0011] - Capacity: Typical operating frequencies in digital electronic platforms are on the order of a few GHz, which represents a difficult limitation to overcome when increasing capacity. Another important limitation lies in the interconnection of the architecture's different subsystems. Conventional electronic systems require sequenced signals to effectively implement the massively parallel interconnections presented by neural networks, which ultimately creates a trade-off between bandwidth and capacity. Additionally, the latency that typically exists in memory read / write processes in digital systems also negatively impacts the total capacity achievable by the system.

[0012] - Scalability: In an attempt to offset current operating frequencies, conventional electronic solutions improve capacity by parallelizing physical resources, enabling the simultaneous processing of large amounts of data. This inevitably translates into the need to increase the use of hardware resources.

[0013] - Power consumption: Modern AI algorithms are typically fed large data sets and required to execute with a high degree of accuracy. Under this scenario, ultra-high capacity is required. Parallelizing hardware resources comes at the cost of increased power consumption.

[0014] Furthermore, high energy consumption is often related to the generation of large thermal loads that must be dissipated. The parallelization of racks containing multiple GPUs or TPUs mounted on servers is the basis of today's large data centers. These centers have strict thermal management requirements for proper operation. In fact, the energy required to cool the facilities and thus ensure proper operation could be as large as the power consumption of the servers themselves.

[0015] The above difficulties can be partially solved by using photonic processors:

[0016] - On the one hand, arithmetic operations, such as MACs, can potentially be performed at the speed of light on photonic platforms and are ultimately limited by the data modulation frequency.

[0017] - On the other hand, optical signals suffer from lower attenuation per unit length and lower latency (below 100 ps), dissipate less heat and are less susceptible to electromagnetic interference, resulting in more compact and efficient solutions with higher performance in terms of capacity per unit area.

[0018] Among the state-of-the-art documents that have developed photonic processors we find:

[0019] - Application US2022012582A1 describes optical neurons using a single wavelength based primarily on coherent linear stages and a non-linear activation and conversion unit.

[0020] Using the optical neuron above, an optical neural network is illustrated in Figure 1. The neural network depicted includes multiple layers of neurons with electronic output signals from one layer being used to determine neuron inputs to the next layer (specifically, the electronic controller circuitry associated with the next layer).

[0021] - On the other hand, application US2022045757A1 describes individual network layers implemented by coherent optical linear neurons in a crossbar configuration, as shown in Figure 2, which are integrated with electronic circuitry at the interfaces between neural network layers to determine neuron inputs to a layer based on neuron outputs from the previous layer. Wavelength division multiplexing (WDM) can be used to efficiently implement certain specific network models, optionally in conjunction with electro-optical switches to make a generic hardware configuration programmable.

[0022] The main difference arises from the degree of coherence of the light source(s). The technique used in the identified patents uses high-coherence (monochromatic) light sources. That is, the phase stability of the light sources must be very high over relatively long periods of time. To achieve this, both the thermal stability and the stability of the light source's power signals are critical. In this context, thermal and power management systems are generally more complex than in systems based on low coherence. Additionally, the techniques used in the patents require multiple detectors, one for each neuron in high-coherence schemes. This results in a more complex solution with less potential in terms of integration density.

[0023] Therefore, the object of the present invention is to overcome the drawbacks of the state of the art based on low-coherence light sources, which results in low optical coherence interferometry. All neurons that are part of a given layer can be multiplexed in the coherence domain (CDM). This fact simplifies the detection scheme and the output of each neuron can be recovered using demultiplexing techniques in the time domain or in the frequency domain with a single detector for each layer, instead of requiring one for each neuron. Consequently, our invention can result in a less complex and more affordable solution with greater potential in terms of integration density. DESCRIPTION OF THE INVENTION

[0024] The object of the present invention is set forth in its essence in the independent claim and the different embodiments are set forth in the dependent claims.

[0025] The present invention aims to design new photonic processor architectures that enable the parallelization of MAC-type arithmetic operations, while maintaining the promise of achieving negligible power consumption during data inference. The basic embodiment implementing a single layer of an ANN is shown in Figure 3.

[0026] The architecture of a neuromorphic photonic processor using low coherence interferometry designed to process a plurality of inputs, comprising the optical artificial neural network layer:

[0027] - a light source configured to generate a low optical coherence radiation beam;

[0028] - a multipath optical interferometer comprising:

[0029] • an optical splitter configured to split the low coherence radiation beam into a plurality of optical paths;

[0030] • a plurality of M neurons where each neuron has N inputs defining a plurality of MxN optical paths or interferometric arms each associated with a neuron input and each comprising: o a first optical amplitude modulator for imparting the corresponding data input to the radiation beam; o a second optical amplitude modulator for imparting the corresponding absolute value of the weight to the radiation beam; o a phase retarder configured to impart a phase delay to the radiation beam associated with the product of the signs corresponding to the data and weight values; and / or an optical waveguide configured to propagate the radiation beam along a path such that differences in the lengths of the optical paths associated with different neurons result in the radiation beams being non-coherent with respect to each other;

[0031] • an optical path comprising a delay line configured to generate a variable optical time delay associated with the maximum length difference between all optical paths;

[0032] • an optical combiner coupled to the outputs of the plurality of optical paths;

[0033] • a detection stage configured to create at least as many optical interference patterns as the number of neurons and comprising: o an electro-optical converter to transform the combined radiation beam into an electrical signal; o an electronic circuit configured to detect the envelope of the electrical signal; and / or a module configured to generate a non-linear activation transfer function

[0034] The light splitting or combining means may be based, for example, on a single or a plurality of multi-mode interference (MMI) waveguides; or on a single or a plurality of directional couplers.

[0035] Each interferometric arm consists of at least two amplitude modulators, a phase retarder, and a custom-length optical path. An amplitude modulator is intended to impart the corresponding absolute value of the data to the input / of neuron j, |X Z ®|. The second amplitude modulator is intended to impart the absolute value of the weight corresponding to the input / of neuron j, |W Z ®|. Both amplitude modulators can be Mach-Zehnder interferometers, but any other amplitude modulation technology could be used. The phase retarder is intended to encode the product of the signs of the input data X z ® and the weight Wz ®. If both amplitude modulators do not generate a residual phase modulation (RPM) accompanying the amplitude modulation process, then a phase delay of TT radians results in destructive interference and therefore a negative overall sign (signal subtraction). Conversely, if zero phase delay is induced on the optical signal, a positive overall signal mediated by constructive interference is obtained (signal addition). In the case of RPM, a complementary phase delay can be induced to reach TT or 0 radians to generate destructive or constructive interference, respectively.

[0036] The K number may ultimately be limited by the maximum group delay generated by the lower arm of the interferometer (reference arm) and the added optical loss.

[0037] After opto-electrical conversion, a low-pass filter is implemented in order to detect the envelope of the interference patterns.

[0038] Because the proposed embodiment is based on low-coherence light sources (e.g., broadband sources), the result is low-coherence interferometry. In this case, all neurons in a given layer can be coherence-domain multiplexed (CDM). This simplifies the detection system, and each neuron's output can be recovered using time-domain or frequency-domain demultiplexing techniques with a single detector (as opposed to a plurality of detectors, one for each neuron in the layer, as required in high-coherence schemes). Consequently, our invention can result in a less complex and more affordable solution with greater potential in terms of integration density.

[0039] Unless otherwise indicated, all technical and scientific elements used herein have the meaning commonly understood by one of ordinary skill in the art to which this invention pertains. Methods and materials similar or equivalent to those described herein can be used in the practice of the present invention.

[0040] Throughout the description and claims, the word "comprises" and its variants are not intended to exclude other technical features, additives, components, or steps. For those skilled in the art, other objects, advantages, and features of the invention will be apparent in part from the description and in part from the practice of the invention.

[0041] EXPLANATION OF THE FIGURES

[0042] To complement the description being made and in order to help better understand the characteristics of the invention, in accordance with a preferred example of practical implementation thereof, a set of drawings is attached as an integral part of said description, in which the following has been represented for illustrative and non-limiting purposes.

[0043] In Figure 1, we can observe a neuromorphic computing scheme comprising linear coherent optical neurons implemented with multipath optical interferometers corresponding to patent US2022045757 A1. In Figure 2, we can observe coherent photonic circuit architectures that optically implement linear algebraic computations exploiting wavelength multiplexing, corresponding to patent US2022012582 A1.

[0044] Figure 3 shows the basic scheme of the neuromorphic photonic processor object of the invention.

[0045] Figure 4 shows the connection between the electronic circuit for detecting the envelope and the nonlinear activation module.

[0046] Figure 5 shows a timeline for the case of a single neuron composed of two inputs, where the absolute value of the weights remains constant during inference while the input data is modulated at a given frequency. For simplicity, the contribution of the signs has been omitted.

[0047] Figure 6 shows a timeline for the case of a layer consisting of four neurons (M=4) and an arbitrary number N of inputs per neuron. At the output of the nonlinear activation (NLA) circuit, a digitizing stage (ADC) is proposed, although its presence will depend on the selected interlayer connection strategy.

[0048] Figure 7 shows a first complementary embodiment in which an additional arm has been added to the interferometer to have extended control over the dynamic range and to be able to represent both positive and negative numbers.

[0049] Figure 8 shows a second embodiment within the principles of the invention in which a reference arm is incorporated with a length equivalent to the optical path of each neuron, instead of being a common element for all neurons that form part of the layer. Figure 9 shows an alternative version of the previous embodiment in which a section of the delay line of the reference arm is common to all neurons, while other sections apply to each neuron individually.

[0050] An alternative version of the previous embodiment is shown in Figure 10 in which a non-linear activation module is inserted at the output of each neuron and before the optical combiner, to the detriment of inserting a non-linear activation module in the detection stage.

[0051] Figure 11 shows a similar embodiment to that of Figure 9, but this time with the interferometric arms grouped according to the data modulators, so that their number is minimized.

[0052] PREFERRED EMBODIMENT OF THE INVENTION

[0053] A preferred embodiment of the proposed invention is described below in view of the figures.

[0054] In Figure 3 we can see the architecture of a neuromorphic photonic processor using low coherence interferometry designed to process a plurality of inputs, comprising the optical artificial neural network layer:

[0055] - a light source (1) configured to generate a low coherence radiation beam

[0056] - a multipath optical interferometer comprising:

[0057] • an optical splitter (2) configured to split the low coherence radiation beam into a plurality of optical paths;

[0058] • a plurality of M neurons (3) where each neuron (3) has N inputs defining a plurality of MxN optical paths or interferometric arms (4) each associated with a neuron input and each comprising: o a first optical amplitude modulator (4.1) for imparting the corresponding data input to the radiation beam; o a second optical amplitude modulator (4.2) for imparting the corresponding absolute value of the weight to the radiation beam; o an optical phase retarder (4.3) configured to impart a phase delay to the radiation beam associated with the product of the signs of both the data and the weight; and o an optical waveguide (4.4) configured to propagate the radiation beam along a path such that differing optical path lengths associated with different neurons result in the radiation beams being non-coherent with each other;

[0059] • an optical path (5) comprising a delay line configured to generate a variable optical time delay associated with the maximum length difference between all optical paths;

[0060] • an optical combiner (6) coupled to the outputs of the plurality of optical paths;

[0061] • a detection stage (7) configured to generate at least as many optical interference patterns as the number of neurons and comprising: o an electro-optical converter (8) to transform the combined radiation beam into an electrical signal; o an electronic circuit (9) configured to detect the envelope of the electrical signal; and o a non-linear activation (NLA) module (10) (Figure 4) configured to generate a non-linear activation transfer function. The light source (1) may be a broadband superluminescent diode (SLED) or a pulsed laser with pulse duration in the femtosecond range.

[0062] Figure 4 shows how the signal at the output of the envelope detector (9) enters the non-linear activation (NLA) module (10) before possible digitization. The NLA function can be different depending on the nature of the input data to be inferred and the type of layer. Typical NLA functions can be sigmoid, hyperbolic tangent, rectified linear (ReLu), Softmax or exponential, to name a few. A ReLu function is considered in Figure 4.

[0063] As an example, consider a simple scenario where the layer consists of a single neuron (M = 1) fed by 2 inputs (N = 2). Considering positive accumulation, that is, = = or, then the output of the optical neuron is calculated as:

[0064] An example schedule is illustrated in Figure 5. The absolute value of the weights remains constant during inference while the input data is modulated at a frequency f c = i / t2- .

[0065] DL corresponds to the reference arm's feed signal, which establishes a variable group delay in order to construct the interferometric pattern. The group delay shift induced by modulating the reference arm must be at least the length of the interferometric pattern (usually taken as several times the coherence length of the light source, typically 4).

[0066] If the previous case is extrapolated to a more complex layer composed of 4 neurons (M = 4) and an arbitrary number of inputs N per neuron, then an example of a schedule is shown in Figure 6. In this case, the master clock frequency is i / t c , where t c = t2- Within the period of the master clock, the four interferometric patterns are adjusted. Therefore, the time spent on each one t M = t c / M must be greater than the duration of each interferometric pattern to avoid temporal overlap (a safe option is to select a length at least 4 times greater than the length of / ) / coherence of the light source) t M > c a / c .

[0067] DL should be configured to establish a variable group delay in order to build the interferometric patterns by sequentially scanning the optical paths from L (1) to the (M)In this case, a sawtooth-type feed signal is considered as an example. However, a triangular-type drive signal can be used to eliminate the rapid transition between the beginning and end of periods. In this scenario, the interferometric patterns would be reversed when scanning the negative slope compared to those constructed when scanning the positive slope.

[0068] The delay line DL in the reference arm needs to sweep from ¿w , which is the physical length of the arm with a shortest optical path, to L (1) + L (M) , being > M ■ 4i c with i c the coherence length of the light source.

[0069] As a result, the photodetector bandwidth depends on the clock frequency of the DL drive signal. Generally speaking, the required detection bandwidth depends on both the number of neurons and the clock frequency, due to the Doppler shift. This can ultimately limit the maximum operating frequency. In this approach, an envelope detector and a nonlinear analog drive circuit are implemented. Analog-to-digital conversion can then take place. The sampling rate can be as low as the clock frequency.

[0070] Figure 7 shows that an additional arm (13) can be added to the interferometer for each neuron to have extended control over the dynamic range at the output of each neuron and the schematic is shown in this embodiment. Consequently, constructive or destructive interference between the weighted sum signal and a constant signal level whose amplitude is preferably fixed at the center of the dynamic range, results in an optical pulse for a positive sum and in a power drop around said fixed amplitude level for a negative sum.

[0071] The additional arm (13) for each neuron (3) is composed, at least, of:

[0072] • an optical amplitude modulator (13.1) configured to impart the corresponding absolute value of the weight to the radiation beam;

[0073] • an optical phase retarder (13.2) configured to impart a phase delay to the radiation beam associated with the weight sign; and

[0074] • an optical waveguide (13.3) configured to propagate the radiation beam along a path such that the difference in length of the optical paths associated with different neurons results in the radiation beams not being coherent with each other.

[0075] To form a neural network composed of a plurality of layers, a plurality of these described architectures can be concatenated in a cascade, so that the ANN with the required dimension is built.

[0076] A second alternative embodiment is shown in Figure 8, in which a reference arm (14) is incorporated in each neuron with a length equivalent to the optical path of each neuron. Consequently, the maximum operating bandwidth would remain constant regardless of the number of neurons M and limited to the length of a single interference pattern and the clock frequency. In the case of implementing a detection scheme based on frequency diversity, each pattern can be located in a different spectral zone by means of controlling the maximum delay induced to each reference arm that is part of each neuron.

[0077] The reference arm (14) in each neuron (3) may comprise: • an optical delay line (14.1) configured to generate a variable time delay associated with the temporal length of a single interference pattern; and

[0078] • an optical waveguide (14.2) configured to propagate the radiation beam along a path such that the difference in length of the optical paths associated with different neurons results in the radiation beams not being coherent with each other.

[0079] An improved version of the previous embodiment with a smaller number of modulation sections is shown in Figure 9. The idea is to share a multipath delay line (15) in the reference arm.

[0080] The multipath delay line (15) comprises:

[0081] • an optical delay line (15.1 ) configured to generate a variable time delay associated with the temporal length of a single interference pattern;

[0082] • an optical splitter (15.2) configured to split the low coherence radiation beam into a plurality of optical paths;

[0083] • a plurality of optical delay lines (15.3) configured to generate a variable time delay associated with the Doppler shift in frequency;

[0084] • a plurality of optical waveguides (15.4) configured to propagate the radiation beams along a plurality of paths, such that the difference in length of the optical paths associated with different neurons results in the radiation beams not being coherent with each other; and

[0085] • an optical combiner (15.5) configured to combine the low coherence radiation beams of the plurality of optical paths;

[0086] An alternative embodiment is shown in Figure 10, in which the non-linear activation (NLA) module (10) (Figure 4) configured to generate a non-linear activation transfer function is arranged at the output of each neuron (3) and before the optical combiner (6) in a number M, equal to the number of neurons (3) that make up the layer, of optical circuits configured to generate a non-linear activation transfer function. This alternative embodiment is equally applicable to the embodiments shown in Figures 3, 7, 8 and 9.

[0087] Figure 11 below shows a similar implementation to the previous one, but this time with the interferometric arms grouped according to the data modulators. The idea is to reduce the number of modulator sections. In this implementation, the number of modulators reaches 2M(N + i) + M + 1, which is significantly lower for layers formed by a high number of neurons compared to the basic approach of the previous case, which amounts to MN + 1.

[0088] Having sufficiently described the nature of the present invention, as well as the manner of putting it into practice, it is noted that, within its essence, it may be put into practice in other embodiments that differ in detail from the one indicated as an example, and to which the protection sought will also be achieved, provided that it does not alter, change or modify its fundamental principle.

Claims

CLAIMS 1.- Neuromorphic photonic processor using low optical coherence interferometry designed to process a plurality of inputs, the photonic processor comprising an optical artificial neural network layer characterized by: - a light source (1) configured to generate a low optical coherence radiation beam; - a multipath optical interferometer comprising: • an optical splitter (2) configured to split the low coherence radiation beam into a plurality of optical paths; • a plurality of M neurons (3) where each neuron (3) has N inputs defining a plurality of MxN optical paths or interferometric arms (4) each associated with a neuron input and each comprising: o a first optical amplitude modulator (4.1) configured to impart the corresponding data input to the radiation beam; o a second optical amplitude modulator (4.2) configured to impart the corresponding absolute value of the weight to the radiation beam; o an optical phase retarder (4.3) configured to impart a phase delay to the radiation beam associated with the product of the signs of both the data and the weight; and o an optical waveguide (4.4) configured to propagate the radiation beam along a path such that differing optical path lengths associated with different neurons result in the radiation beams being non-coherent with respect to each other; • an optical path (5) comprising an optical delay line configured to generate an associated variable time delay with the maximum length difference between all optical paths; • an optical combiner (6) coupled to the outputs of the plurality of optical paths; • a detection stage (7) configured to generate at least as many optical interference patterns as the number of neurons and comprising: o an electro-optical converter (8) to transform the combined radiation beam into an electrical signal; o an electronic circuit (9) configured to detect the envelope of the optical signal; and / or a non-linear activation (NLA) module (10) configured to generate a non-linear activation transfer function. 2.- Neuromorphic photonic processor using low optical coherence interferometry according to claim 1 characterized in that the light source (1) is a broadband superluminescent diode (SLED). 3.- Neuromorphic photonic processor using low optical coherence interferometry according to claim 1, characterized in that the light source (1) is a pulsed laser with pulse duration in the femtosecond range. 4.- Neuromorphic photonic processor using low optical coherence interferometry according to any of claims 1, 2 or 3, characterized in that an additional arm (13) is added to each of the neurons (3) configured to provide greater control over the dynamic range at the output of each neuron (3). 5.- Neuromorphic photonic processor using low optical coherence interferometry according to claim 4, characterized by an additional arm (13) for each neuron (3) composed of at least: • an optical amplitude modulator (13.1) configured to impart the corresponding absolute value of the weight to the radiation beam; • an optical phase retarder (13.2) configured to impart a phase delay to the radiation beam associated with the weight sign; and • an optical waveguide (13.3) configured to propagate the radiation beam along a path such that the difference in length of the optical paths associated with different neurons results in the radiation beams not being coherent with each other. 6.- Neuromorphic photonic processor by means of low optical coherence interferometry according to any of claims 1, 2, 3, 4 or 5, characterized in that the optical path (5) comprises a plurality M of optical paths (14) inserted in each neuron (3), where an optical path (14) comprises: • an optical delay line (14.1 ) configured to generate a variable time delay associated with the temporal length of a single interference pattern; and • an optical waveguide (14.2) configured to propagate the radiation beam along a path such that the difference in length of the optical paths associated with different neurons results in the radiation beams not being coherent with each other. 7.- Neuromorphic photonic processor by means of low optical coherence interferometry according to any of claims 1, 2, 3, 4 or 5, characterized in that the optical path (5) comprises a multipath delay line (15) comprising: • an optical delay line (15.1 ) configured to generate a variable time delay associated with the temporal length of a single interference pattern; • an optical splitter (15.2) configured to split the low coherence radiation beam into a plurality of optical paths; • a plurality of optical delay lines (15.3) configured to generate a variable time delay associated with the Doppler shift in frequency; • a plurality of optical waveguides (15.4) configured to propagate the radiation beams along a plurality of paths, such that the difference in length of the optical paths associated with different neurons results in the radiation beams not being coherent with each other; and • an optical combiner (15.5) configured to combine the low coherence radiation beams of the plurality of optical paths; 8.- Neuromorphic photonic processor by means of low optical coherence interferometry according to any of claims 1, 2, 3, 4, 5, 6 or 7, characterized in that the non-linear activation (NLA) module (10) configured to generate a non-linear activation transfer function comprises a plurality M of optical circuits inserted at the output of the neurons (3) and before the optical combiner (6). 9.- Neuromorphic photonic processor by means of low optical coherence interferometry according to any of claims 1, 2, 3, 4, 5, 6 or 7, characterized in that the first optical amplitude modulators (4.1) configured to impart the corresponding data input to the radiation beams for each neuron (3) are arranged in a common manner to interference arms (4) corresponding to different neurons (3), so that the number of these elements can be reduced independently of the number of neurons (3).

Citation Information

Patent Citations

  • Neuromorphic photonics with coherent linear neurons

    US20220012582A1

  • Coherent photonic computing architectures

    US20220045757A1

  • Integrated delay line for optical coherence tomography

    ES2396391A2

  • Scanning device for low coherence interferometry

    US20140078510A1

  • Optical processing

    WO2000072107A1