Neural network method and apparatus with duplication, parallel processing, and combination of signals

Abstract

A method of operating a neural network, comprising: duplicating a signal to be input into a computational layer of the neural network, in order to generate at least two copies of the signal; processing, in parallel, the at least two copies of the signal, using at least two processing components configured to perform at least equivalent processing operations; and combining the at least two processed copies of the signal, to generate at least one, preferably exactly one, output signal.

Description

[0001] Neural network method and apparatus with duplication, parallel processing, and combination of signals

[0002] TECHNICAL FIELD

[0003] The present disclosure generally relates to neural networks, and in particular to a method of operating a neural network and to a data processing apparatus storing a neural network.

[0004] BACKGROUND

[0005] Neuromorphic photonics has shown promise in combining photonics and neural network computing. Photonic neural networks leverage light's high-speed transmission and parallel processing for simulating neural behaviour. Prior art has introduced photonic neuron models, reconfigurable synapses, on-chip signal processing, and optical signal multiplexing to enhance network capabilities. Despite progress, challenges persist in achieving higher performance, scalability, and energy efficiency.

[0006] SUMMARY

[0007] It is an aim of at least some embodiments according to the present disclosure to improve accuracy of computation in neural networks. It is an aim of at least some particular embodiments to do this for photonic neural networks.

[0008] Accordingly, in a first aspect of the present disclosure, there is provided a method of operating a neural network; the method comprising:

[0009] - duplicating a signal to be input into a computational layer of the neural network, in order to generate at least two copies of the signal;

[0010] - processing, in parallel, the at least two copies of the signal, using at least two processing components configured to perform at least equivalent processing operations; and

[0011] - combining the at least two processed copies of the signal, to generate at least one, preferably exactly one, output signal. By duplicating the signal into multiple and then processing the multiple copies in parallel, thus independently, using sufficiently similar downstream processing components, the neural network can be operated in a manner that improves accuracy, because any noise that might creep into any one or more of the copies is likely to be balanced out by the existence of the other copies (which may themselves even also be subject to noise). By combining the outcomes of these parallel processes, the harmful impact of noise on individual copies may be alleviated or mitigated. This can help to improve the overall accuracy of the neural network.

[0012] In this context, the term duplicating may be taken to refer to an act of making a copy of something, in particular an identical or at least corresponding copy.

[0013] In other words, by expanding the number of copies of a signal, then processing the separate copies, and by then contracting or collapsing the number of copies back to one or fewer copies, a benefit relating to the law of large numbers or wisdom of the crowd may be introduced into the operation of the neural network. In other words, this allows the neural network to balance complexity, cost and energy use on the one hand for improved accuracy on the other hand, by a form of duplication.

[0014] In various embodiments, the combining comprises reducing an impact of noise introduced into at least one copy of the at least two copies on the at least one output signal.

[0015] In other words, if there is noise that interfered with a copy during its generation in the duplicating or during its processing, the combining may allow to mitigate or suppress the impact of this noise, for example thanks to the contribution of other copies which may be subject to no or less or different noise.

[0016] In various embodiments, the neural network is an optical neural network, and wherein the duplicating comprises optical splitting of the signal.

[0017] In other words, the neural network may preferably be a photonic neural network (whereas in other embodiments, a non-photonic neural network may be considered). Using this embodiment of the method for a photonic neural network has the advantage that light signals in the photonic neural network may be easily combined with each other via superposition, potentially requiring less overhead compared to non-photonic neural networks.

[0018] In various embodiments, the optical splitting of the signal comprises splitting the signal into at least two copies of the signal, and further comprises wavelength conversion wherein the at least two copies have at least two different wavelengths.

[0019] Advantageously, this allows the separate copies to utilize the same hardware partially or entirely, given that light of different wavelengths does not interfere with itself.

[0020] In various embodiments, the combining comprises reducing at least one of an amplitude and a phase of the at least two copies of the signal, in comparison to a sum of the at least two copies of the signal, such that a result of the combining remains bounded by acceptance specifications of at least a subsequent part of the neural network.

[0021] Advantageously, this helps to ensure that the resultant output signal does not grow too large or to a phase outside of the bounds of the rest of the neural network, which helps to prevent inaccuracies or even failure of the neural network.

[0022] In various embodiments, the combining of the at least two processed copies of the signal comprises averaging of the at least two processed copies of the signal.

[0023] Advantageously, this technique is easy and straightforward, and mathematically effective to combine the copies. Of course, different techniques may be considered instead or additionally.

[0024] In various embodiments, the method comprises repeating the steps of: the duplicating, the processing, and the outputting; wherein the repeating is preferably performed such a number of times, that a final output signal is as error-free as a corresponding output signal of a corresponding noiseless neural network. In this manner, the neural network may be made arbitrarily layered, wherein sequential units, each unit including a step of duplication, processing and combination, may follow each other. Moreover, such units may also operate in parallel to each other (as will be described below with reference to Figures 1 and 2).

[0025] Moreover, in a second aspect of the present disclosure, there is provided a data processing apparatus storing a neural network, wherein the neural network comprises at least one duplicating means configured for duplicating a signal to be input into a computational layer of the neural network, in order to generate at least two copies of the signal; wherein the neural network comprises at least two processing components configured to perform at least equivalent processing operations, in parallel, on the at least two copies of the signal; and wherein the neural network comprises at least one combining means configured for combining the at least two processed copies of the signal, to generate at least one, preferably exactly one, output signal.

[0026] In various embodiments, the at least one combining means is configured to reduce an impact of noise introduced into at least one copy of the at least two copies on the at least one output signal.

[0027] In various embodiments, the neural network is an optical neural network, and wherein the at least one duplicating means comprises an optical splitter configured for optically splitting the signal.

[0028] In various embodiments, the optical splitter is a wavelength conversion splitter and is configured for splitting the signal into at least two copies of the signal, wherein the at least two copies have at least two different wavelengths.

[0029] In various embodiments, the at least one combining means is configured for reducing at least one of an amplitude and a phase of the at least two copies of the signal, in comparison to a sum of the at least two copies of the signal, such that a result of the combining remains bounded by acceptance specifications of at least a subsequent part of the neural network.

[0030] In various embodiments, the combining of the at least two processed copies of the signal comprises averaging of the at least two processed copies of the signal.

[0031] In various embodiments, the data processing apparatus comprises a plurality of duplicating means, a plurality of processing components, and a plurality of combining means, defining a plurality of units, each unit comprising a duplicating means, at least two processing components, and a combining means, and preferably arranged such that a final output signal is as error-free as a corresponding output signal of a corresponding noiseless neural network.

[0032] The embodiments described herein are provided for illustrative purposes and should not be construed as limiting the scope of the invention. It is to be understood that the invention encompasses other embodiments and variations that are within the scope of the appended claims. The invention is not restricted to the specific configurations, arrangements, and features described herein. The invention has wide applicability and should not be limited to the specific examples provided. The embodiments disclosed are merely exemplary, and the skilled person will appreciate that various modifications and alternative designs can be made without departing from the scope of the invention.

[0033] DESCRIPTION

[0034] The above-described method and data processing apparatus will be more fully understood with the help of the detailed description below as well as the appended description documents, whose contents are included within the present description in their entirety. The skilled person will understand that those included contents are intended as an exemplary description relating to various embodiments of the method and the data processing apparatus according to the present disclosure, and are not meant to be interpreted in a limiting manner. In particular, various features and elements, taken individually or in combination, of the embodiments described therein may be considered optional in view of the embodiments described above and the embodiments represented by the appended claims.

[0035] The problem of noise-accumulation in photonic computation is combatted, especially suitable for Optical Neural Network application.

[0036] The goal of achieving high-accuracy in photonic computing is tackled through physically unaware training out of the photonic chip and calibration.

[0037] Analog data processing is fundamentally subject to noise. The noise could be combatted through executing the same calculations multiple times but in parallel and averaging their results. We propose a design / an arrangement of optical components such that noise-reduction is achieved via copying and averaging calculations.

[0038] Signals within an optical neural network are repeatedly split, possibly using multiple wavelengths, so as to duplicate the carried information at intermediate stages. Each duplicated piece of information is passed through an independent, but identical, optical subcomponent. These distinct but similarly processed pieces of information are recombined deeper within the optical neural network so as to create optical signals that are more robust to component noise.

[0039] Two possible implementations are detailed in the paper, that is attached to this document. Schematic representations of the two designs in the paper are given in Figure 1 and Figure 2.

[0040] Figure 1 represents an accordion-like design: the light colored circles 103 represent weighted addition, i.e. , matrix-vector products. The dark colored circles 103 stand for averaging between the results of the weighted-addition. Boxes 105 represent the activation function and thus the output of each layer.

[0041] Figure 2 represents a tree-like design: again, the light colored circles 201 represent weighted addition, i.e., matrix-vector products. The dark colored circles 203 stand for averaging between the results of the weighted-addition. Boxes 205 represent the activation function and thus the output of each layer.

[0042] Of course, a combination of the accordion-like and the tree-like design may also be considered, wherein the accordion-like arrangement is combined with and / or alternated with the tree-like arrangement of the other. Furthermore, different designs may be used as well for practical implementations.

[0043] An exemplary way to implement the method will now be described: Consider the chip presented in:

[0044] Shi, B., Calabretta, N. and Stabile, R., 2019. Deep neural network through an InP SOA-based photonic integrated cross-connect. IEEE Journal of Selected Topics in Quantum Electronics, 26(1), pp.1-11.

[0045] Figure 3 schematically illustrates a chip from Shi et al., which is suitable for an implementation of the proposed design.

[0046] The design can be implemented by using the same data x on the left hand side of the image. Then the outputs (y_1 , ... , y_8) are noisy versions of the true value. Joining and splitting the signals y_1 , ... , y_8 will create new signals y_1 , ... , y_8, which are all the same, also noisy, but with lower variance and distributed around the true value.

[0047] The presently described embodiments may be modified to suit different ONN platforms and layouts.

[0048] Optical Neural Networks are fast and power efficient in inference, but pose challenges because of their intrinsic noise. Therefore, embodiments according to the present disclosure may be especially practical for applications benefiting from high-accuracy inference at high speed, including real-time Al applications, for example robotics and autonomous driving, or other high-speed applications, like for example error corrections in optical transmission, and in finance. Photonic neural networks are attracting more and more interest, and a near future product development of those for various applications, from image classification to Ising machines to green optical communication, may be foreseen.

[0049] It is moreover noted that embodiments according to the present disclosure are more or less independent from the precise type of data used to train the neural network or to make inferences with it. These embodiments rather relate to the organisation and use of such neural networks. Therefore, the present disclosure rather relates to computational and physical organisation of neural networks. In that sense, the present disclosure may be applicable to any type of neural network, especially when there is a risk of noise entering the neural network.

[0050] As used in this application and in the claims, the singular forms “a,” “an,” and “the” include the plural forms unless the context clearly dictates otherwise. The systems, apparatus, and methods described herein should not be construed as limiting in any way. Instead, the present disclosure is directed toward all novel and non-obvious features and aspects of the various disclosed embodiments, alone and in various combinations and sub-combinations with one another. The disclosed systems, methods, and apparatus are not limited to any specific aspect or feature or combinations thereof, nor do the disclosed systems, methods, and apparatus require that any one or more specific advantages be present or problems be solved. Any theories of operation are to facilitate explanation, but the disclosed systems, methods, and apparatus are not limited to such theories of operation.

[0051] Although the operations of some of the disclosed methods are described in a particular, sequential order for convenient presentation, it should be understood that this manner of description encompasses rearrangement, unless a particular ordering is required by specific language set forth below. For example, operations described sequentially may in some cases be rearranged or performed concurrently. Moreover, for the sake of simplicity, the attached figures may not show the various ways in which the disclosed systems, methods, and apparatus can be used in conjunction with other systems, methods, and apparatus. Additionally, the description sometimes uses terms like “obtaining” and “outputting” to describe the disclosed methods. These terms are high-level abstractions of the actual operations that are performed. The actual operations that correspond to these terms will vary depending on the particular implementation and are readily discernible by the skilled person.

[0052] It will be appreciated that numerous specific details are set forth in order to provide a thorough understanding of the examples described herein. However, it will be understood by the skilled person that the examples described herein can be practiced without these specific details. In other instances, methods, procedures and components have not been described in detail so as not to obscure the related relevant feature being described. The drawings are not necessarily to scale and the proportions of certain parts may be exaggerated to better illustrate details and features. The description is not to be considered as limiting the scope of the examples described herein.

Claims

CLAIMS1. A method of operating a neural network; the method comprising:- duplicating a signal to be input into a computational layer of the neural network, to generate at least two copies of the signal;- processing, in parallel, the at least two copies of the signal, using at least two processing components configured to perform at least equivalent processing operations; and- combining the at least two processed copies of the signal, to generate at least one, preferably exactly one, output signal.

2. The method of claim 1 , wherein the combining comprises reducing an impact of noise introduced into at least one copy of the at least two copies on the at least one output signal.

3. The method of any preceding claim, wherein the neural network is an optical neural network, and wherein the duplicating comprises optical splitting of the signal.

4. The method of claim 3, wherein the optical splitting of the signal comprises splitting the signal into at least two copies of the signal, and further comprises wavelength conversion wherein the at least two copies have at least two different wavelengths.

5. The method of any preceding claim, wherein the combining comprises reducing at least one of an amplitude and a phase of the at least two copies of the signal, in comparison to a sum of the at least two copies of the signal, such that a result of the combining remains bounded by acceptance specifications of at least a subsequent part of the neural network.

6. The method of any preceding claim, wherein the combining of the at least two processed copies of the signal comprises averaging of the at least two processed copies of the signal.

7. The method of any preceding claim, comprising repeating the steps of: the duplicating, the processing, and the outputting; wherein the repeating is preferably performed a number of times such that a final output signal is as error-free as a corresponding output signal of a corresponding noiseless neural network.

8. A data processing apparatus storing a neural network, wherein the neural network comprises at least one duplicating means configured for duplicating a signal to be input into a computational layer of the neural network, in order to generate at least two copies of the signal; wherein the neural network comprises at least two processing components configured to perform at least equivalent processing operations, in parallel, on the at least two copies of the signal; and wherein the neural network comprises at least one combining means configured for combining the at least two processed copies of the signal, to generate at least one, preferably exactly one, output signal.

9. The data processing apparatus of claim 8, wherein the at least one combining means is configured to reduce an impact of noise introduced into at least one copy of the at least two copies on the at least one output signal.

10. The data processing apparatus of any preceding claim, wherein the neural network is an optical neural network, and wherein the at least one duplicating means comprises an optical splitter configured for optically splitting the signal.

11. The data processing apparatus of claim 10, wherein the optical splitter is a wavelength conversion splitter and is configured for splitting the signal into at least two copies of the signal, wherein the at least two copies have at least two different wavelengths.

12. The data processing apparatus of any of claims 8-11 , wherein the at least one combining means is configured for reducing at least one of an amplitude and a phase of the at least two copies of the signal, in comparison to a sum of the at least twocopies of the signal, such that a result of the combining remains bounded by acceptance specifications of at least a subsequent part of the neural network.

13. The data processing apparatus of any of claims 8-12, wherein the combining of the at least two processed copies of the signal comprises averaging of the at least two processed copies of the signal.

14. The data processing apparatus of any of claims 8-13, comprising a plurality of duplicating means, a plurality of processing components, and a plurality of combining means, defining a plurality of units, each unit comprising a duplicating means, at least two processing components, and a combining means, and preferably arranged such that a final output signal is as error-free as a corresponding output signal of a corresponding noiseless neural network.

Images