Optical image processing

Abstract

The device (2) comprises a medium (7) capable of supporting overlapping optical modes (10) such that, in response to spatially-structured illumination (4), a set of overlapping optical modes (10) in t

Description

Field of the Invention The present invention relates to optical image processing. Background A promising approach to machine learning is to use physical neuromorphic computing schemes. In physical neuromorphic computing schemes, hardware architectures are used to emulate the function of neurons and / or neural networks. Various approaches have been made to implement physical neuromorphic computing schemes; however, these can have one or more drawbacks. Memristor arrays can suffer from poor dimensionality, lack a capability for image input, and rapidly degrade physically. Optical delay-line systems can involve reels of kilometres of optical cable and thereby be incompatible with miniaturisation and being employed in remote non-data-centre use cases such as in self-driving cars. Vertical emitting cavity lasers ("VCSELs") can lack large input dimensionality capability and not allow for mode separation. Moreover, the data input in many existing neuromorphic computing schemes is typically one-dimensional or few-dimensional. This can be a drawback because many modern machine learning problems, such as those relating to computer vision, concern high-dimensional data. Physical neuromorphic computing schemes have been Implemented in a range of all-optical and optoelectronic systems. T. Zhou et al.: "Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit", Nature Photonics, volume 15, page 367 (2021) describes an optoelectronic fused computing architecture based on the diffraction of light. D. Pierangeli, G. Marcucci, and C. Conti: "Photonic extreme learning machine by free-space optical propagation", Photonics Research, volume 9, page 1446 (2021) describes a neuromorphic photonic scheme which can be implemented by using an optical encoder and coherent wave propagation in free space. US 2021 / 0285819 Al describes a system for machine learning using optical data. The system Includes a diffusive physical medium for scattering light from an optical source coupled to an adjustable spatial light modulator. 5 Reference is also made to M. Matuszewski et al. : "Energy-Efficient Neural Network Inference with Microcavity Exciton Polaritons", Physical Review Applied, volume 16, page 024045 (2021), which proposes all-optical neural networks having an energy efficiency of energy efficiency of 4xl016 synaptic operations per second per watt and a performance density 1016 synaptic operations per second per millimetre squared. 10 Software-based approaches can also have one or more drawbacks. A part of software machine learning is the so-called "non-linear activation" step, where data passed through the software neural network undergoes some non-linear transform to add processing power, typically a rectified linear activation function (hereinafter referred to 15 as "ReLU") or a sigmoid transform. Manually engineering software networks so that these transforms can occur in a parallel fashion can have a substantial cost. Summary According to a first aspect of the present invention there is provided a device for use in optical image processing. The device comprises a medium capable of supporting overlapping optical modes such that, In response to spatially-structured illumination, a set of overlapping optical modes in the medium are non-uniformly excited and output a spatially-structured response. This arrangement can allow the device to process data included in the spatially-structured illumination in such a way that the features of the data are conserved. This is because there is no need pre-process the data, for example by pixel rastering (in other words, ID input of pixels as a time series) or by image transforms, before providing it to the device for processing. The medium may be disposed on a surface. The surface may be a surface of the substrate. The surface may be flat. The surface may be non-flat, for example, it may be patterned, curved, or have a non-uniform surface. The device may include the substrate. The medium may be disposed in a matrix of another material. The substrate may be rigid. The substrate may be flexible. The substrate may consist of or consist predominantly of a plastic material. The substrate may be planar. The substrate may be non-planar, for example curved. The substrate may include, consist of, or consist predominantly of, or may include a layer (for example, a surface layer) which includes, consists of, or consists predominantly of a dielectric or a semiconductor, such as indium tin oxide (ITO), silicon dioxide (SiO2), hafnium dioxide (HfO2), titanium dioxide (TiO2), zinc oxide (ZnO), magnesium fluoride (MgF2), or diamond. The substrate may be an oxide material. The substrate may include, consist of, or consist predominantly of a semiconductor material such as silicon (Si). The substrate may include a transparent material and / or a layer (for example, a surface layer) of transparent material, such as a layer of a dielectric or a semiconductor, such as ITO, SIO2, HfO2, TiO2, ZnO, MgF2, or diamond. The medium may comprise an optical gain medium. The optical gain medium may be embedded in the medium. Alternatively, the medium may be an optical gain medium. The optical gain medium may be a solid-state gain medium. The solid state-gain medium may comprise semiconductor material embedded in a polymer material or dye-molecules embedded in a polymer material. Alternatively, the optical gain medium may be a liquid-phase dye medium. The medium may have an optical gain which is less than 1, equal to 1, or greater than 1. In other words, the device may be for amplification of incident light and / or for attenuation of incident light. The medium may be spatially non-uniform. The medium may be spatially non-uniform along at least one in-plane direction. The medium may have in-plane patterning. The medium may be spatially non-uniform along an out-of-plane direction. For example, the medium may be a multi-layered heterostructure. The medium may include voids. The medium may include suspended structures. The medium may be 3D patterned. The medium may be spatially uniform. The medium may be spatially uniform along at least one in-plane direction. The medium may have a complex refractive index that, In response to a mask taking the form of a spatially-structured light beam being projected onto the medium, is spatially structured. The mask may be provided by a second light source that is different to a light source configured to provide the spatially-structured illumination. The medium may be configured as a panel. The medium may be spatially non-uniform across the panel. The medium may be spatially uniform across the panel. The medium may be configured as a screen. The medium may be spatially non-uniform across the screen. The medium may be spatially uniform across the screen. The device may be configured for operation at a design wavelength. The design wavelength may be between 10 nm and 400 nm (or "ultraviolet light"). The design wavelength may be between 400 nm and 780 nm (or "visible light"). The design wavelength may be between 780 nm and 1.4 pm (or "IR-A"). The design wavelength may be between 1.4 pm and 3 pm (or"IR-B"). The design wavelength may be between 3 pm and 1 mm (or "IR-C"). The medium may comprise a first photonic network, the first photonic network comprising network links that are radiatively coupled at nodes. Expressed differently, the medium may comprise a first photonic material and have a first topology in which network links are radiatively coupled at nodes. At least a first sub-set of the network links of the first photonic network may be interconnected at the nodes of the first photonic network. At least a second sub-set of the network links of the first photonic network may be separated by a distance less than the design wavelength at the nodes of the first photonic network. The first photonic network may comprise at least two closed loop paths. The at least two closed loop paths may be for transmitting light having the design wavelength. The medium may comprise a second photonic network, the second photonic network comprising network links that are radiatively coupled at nodes, the second photonic network being spatially separated from the first photonic network. The second photonic network may comprise a second photonic material and have a second topology in which network links are radiatively coupled at nodes. At least a first sub-set of the network links of the second photonic network may be interconnected at the nodes of the second photonic network. At least a second sub-set of the network links of the second photonic network may be separated by a distance less than the design wavelength at the nodes of the second photonic network. The second photonic network may comprise at least two closed loop paths. The at least two closed loop paths may be for transmitting light having the design wavelength. The first photonic material may be different to the second photonic material. The first photonic material may be the same as the second photonic material. The first topology may be different to the second topology. The first topology may be the same as to the second topology. The medium may comprise three or more spatially separated photonic networks. The medium may comprise photonic waveguides. The network links may comprise the photonic waveguides. The photonic waveguides may be embedded in the network links. The network links may be photonic waveguides. The photonic waveguides may be active photonic waveguides. The network links and / or photonic waveguides may have at least one dimension less than 1000 nm. The at least one dimension may include or be a width or a diameter. The at least one dimension may include or be a thickness. The network links and / or photonic waveguides may have at least one dimension greater than or equal to a tenth of the design wavelength. The network links and / or photonic waveguides may have at least one dimension less than or equal to 10 times the design wavelength. The network links and / or photonic waveguides may have at least one dimension less than or equal to 100 times the design wavelength. The network links and / or photonic waveguides may have a distribution of lengths. The distribution of lengths may be centred around an average length value. In the case that the design wavelength is between 400 nm and 780 nm (or "visible light"), and / or in the case that the design wavelength is between 780 nm and 1.4 pm (or "IR-A"), the average length value may be between 3 pm and 20 pm. The average length value may be less than 100 pm. The average length value may be less than 1 mm. The photonic waveguides may have a distribution of diameters. The distribution of diameters may be centred around an average diameter value. The photonic waveguides may have an average diameter value between 1 pm and 10 pm and the average length value may be between 100 pm and 500 pm. The medium may comprise low-dimensional structures. The low-dimensional structures may comprise quantum wells. The quantum wells may be encapsulated quantum wells. The medium may comprise a semiconductor material. The semiconductor material may be an inorganic semiconductor material. The semiconductor material may be a III-V semiconductor material. The III-V semiconductor material may be a direct bandgap III-V semiconductor material. The III-V semiconductor material may be indium phosphide (InP), gallium arsenide (GaAs), or indium gallium arsenide (InGaAs). The semiconductor material may be an organic semiconductor material. The organic semiconductor material may be Rhodamine 6G dye. The organic semiconductor material may be Rhodamine B dye. The set of overlapping optical modes may include optical modes which are strongly coupled. The set of overlapping optical modes may be strongly coupled. The set of overlapping optical modes may include at least 10 optical modes. The set of optical modes may include up to 100 or up to 1000 optical modes. The set of overlapping optical modes may include two or more optical modes that are lasing modes. The set of overlapping optical modes may comprise only optical modes which are not of optical modes which are not lasing modes. The set of overlapping optical modes may comprise or consist of only one optical mode that varies non-linearly with the intensity of the spatially-structured illumination. The spatially-structured illumination may include input data such as an image for classification. The input data may be encoded. The image may be a 2D image. The image may a pixellated image. The image may have a size of at least 20 pixels by 20 pixels, such as 28 pixels by 28 pixels. The image may have a size of at least 1000 pixels by 1000 pixels. The image may be projected optically onto the medium in a native pixel-map state. In the native pixel-map state, local information on neighbouring pixels and image features may be fully preserved. The device may be neuromorphic computing hardware. The device may be for use in neuromorphic optical computing. The device may be capable of providing a test accuracy score, measured by the Modified National Institute of Standards and Technology ("MNIST") handwritten digit classification task, equal to or greater than 95%. The device may be capable of providing a test accuracy score, measured by the MNIST handwritten digit classification task, equal to or greater than 97.7%. The device may be capable of providing a test accuracy score, measured by the MNIST handwritten digit classification task, equal to or greater than 98.15%. According to a second aspect of the present invention there is provided a method of fabricating the device of the first aspect. The method includes: providing a first substrate having a principal surface; providing a hetero-epitaxial wafer having an intermediate sacrificial layer interposed between a second substrate and an upper layer; bonding the upper surface of the hetero-epitaxial wafer to the principal surface of the first substrate; forming a bonded medium, the bonded medium comprising the first substrate and the upper layer, forming the bonded medium comprising etching the sacrificial layer to remove the second substrate and the sacrificial layer; and lithographically patterning structures on the surface of the bonded medium from which the sacrificial layer was etched. The method of the second aspect may include features corresponding to any features of the device of the first aspect. According to a third aspect of the present invention there is provided an optical image processing system. The optical image processing system comprises the device of the first aspect. The optical image processing system also comprises an imaging spectrometer configured to detect the spatially-structured response. Readout from the optical image processing system may be performed spatially uniformly (that is, with one single spatially integrated spectra for each input image for the entirety of the medium) or with spatial resolution (that is, with a spatially integrated spectra for each input image for each of a plurality of defined spatial-sub-regions of the medium). The imaging spectrometer may be configured to perform readout on defined spatial sub-regions of the medium. Readout may be performed on defined spatial sub-regions of the medium using an arrangement of lenses in a detection path between the medium and the spectrometer. Alternatively, readout may be performed on the whole medium. The optical image processing system may further comprise a neutral density filter configured to maintain a pump power of the spatially-structured illumination. The optical image processing system may further comprise a digital micromirror device configured to provide the spatially-structured illumination. The digital micromirror device may be arranged in a reflection geometry. The system may have a configuration in which a majority of functional elements, for example all functional elements, including the medium are arranged in an on-chip package in which functional optical elements are assembled in a multi-layer vertical stack on a flat substrate such as a silicon (Si) chip or the like. The on-chip package (hereinafter referred to as the "photonics chip") may be a flat on-chip package. In the on-chip package, the source of spatially structured illumination may include a spatially uniform flat form factor light source such as a microstrip laser. The optical image processing system may further comprise a spatial light modulator configured to provide the spatially-structured illumination. The spatial light modulator may be used to impart a spatial structure on the light output by a source of illumination. Alternatively, the spatially structured illumination may be provided by an optical pump capable of outputting spatially structured illumination, such as a nanolaser array, a vertical-cavity surface-emitting laser array, or a miniature light-emitting diode array. Readout hardware may include an on-chlp spectrometer of the photonics chip. The photonics chip may further include the medium and the source of illumination. The source of illumination may include photodiodes. The optical Image processing system may be configured for use as a non-trainable reservoir for data processing. The optical image processing system may be configured for use as a fully trainable deep-neural network for inference. The system of the third aspect may include features corresponding to any features of the device of the first aspect and / or the method of the second aspect. According to a fourth aspect of the present invention there is provided a computer vision product. The computer vision product comprises the optical image processing system of the third aspect. Expressed differently, the computer vision product comprises the device of the first aspect and an imaging spectrometer configured to detect the spatially-structured response. The computer vision product also comprises a processor configured to determine variations in spectra provided by the spatially-resolved spectrometer. The spectra may be spatially-resolved spectra. The computer vision product may be selected from the group consisting of a selfdriving car; an autonomous drone, an autonomous aircraft, an autonomous sea vessel, a low-latency image processing system, a CCTV camera processing system, an industrial process control system, a medical imaging system, and a surgical control system. The CCTV camera processing system may be for crowd recognition. The industrial process control system may be for observing a production line, such as a production line for food or a production line for components. The medical imaging system may be for processing data comprising MRI images or tumour images. The surgical control system may be for processing endoscope camera feeds. The computer vision product may be a vehicle capable of being autonomously piloted or a system for object detection and classification. The vehicle may be the self-driving car, the autonomous drone, the autonomous aircraft, or the autonomous sea vessel. The object may be a human face. The system may be the low-latency image-processing system, the industrial process control system, the medical imaging system, or the surgical control system. The computer vision product of the fourth aspect may include features corresponding to any features of the device of the first aspect, the method of the second aspect, and / or the system of the third aspect. According to a fifth aspect of the present invention there is provided a method of operating the device of the first aspect, the optical image processing system of the third aspect, or the computer vision product of the fourth aspect. The method comprises providing the spatially-structured illumination to the device. The method also comprises detecting the spatially-structured response. The spatially-structured response may correspond to a computational task. The spatially-structured response may be reconfigured to correspond to another computational task by training the medium. The method of operating the device may be a method of optical image processing. The method of the fifth aspect may include features corresponding to any features of the device of the first aspect, the method of the second aspect, the system of the third aspect, and / or the computer vision product of the fourth aspect. According to a sixth aspect of the present invention there is provided a method of training the device of the first aspect, the optical image processing system of the third aspect, or the computer vision product of the fourth aspect. The method comprises: determining a loss function of the device, determining whether the loss function has reached a stable minima, and modulating an optical response in the device. The method may comprise iteratively: determining a loss function of the device, determining whether the loss function has reached a stable minima, and modulating an optical response in the device. The optical response may be iteratively modulated using a gradient-descent algorithm or a backpropagation algorithm. The method may be carried out until a determination that the loss function has reached a stable minimum is made. Determining a loss function of the device may comprise assessing task performance for a given task and generating the loss function based on the task performance. Modulating the optical response may include providing interference effects in the medium. The optical response may be a lasing response. The interaction of the spatially-structured illumination with the medium through stimulated emission processes and through photoluminescence may be modulated via application of a field to the medium. Modulating an optical response in the device may comprise adjusting a field applied to the medium. The field may be a local field or a global field. The field be an electric field, a magnetic field, illumination that is different to the spatially-structured illumination, or a temperature field. The field may Include at least two of the group consisting of an electric field, a magnetic field, illumination that is different to the spatially-structured illumination, and a temperature field. Reconfigurable training may be provided by modulating the interaction of the spatially-structured illumination with the medium via application of the field to the medium. Modulating the optical response may Include projecting the mask taking the form of the spatially-structured light beam from the second light source onto at least a part of the medium. Modulating the optical response may include electrically gating the entire medium or locally electrically gating at least one defined spatial sub-region of the medium. A set of electrodes may be used to apply the local electrical gating. The set of electrodes may be patterned laterally adjacent to the medium, above the medium, or below the medium. The set of electrodes may have a 2D arrangement and be configured to apply a locally reconfigurable electric field across the medium. The 2D arrangement may define a 2D grid. The reconfigurable electric field may be used to modulate the optical response. The reconfigurable electric field may be used to modulate a lasing threshold of the medium. Modulating the optical response may include at least one of local magnetic control, local magneto-plasmonic control, global magnetic control, and global magneto-plasmonic control. Magnetic elements such as nanomagnets may be used to provide the local magnetic control or local magneto-plasmonic control. The magnetic elements may be configured to apply a locally reconfigurable magnetic field across the medium. The magnetic state of the magnetic elements may locally control how much of the spatially-structured illumination is absorbed or enhanced and can be used to modulate the optical response. The magnetic state may be reconfigurably programmed by at least one of an applied magnetic field, current / voltage control, or all-optical magnetic switching. Modulating the optical response may include varying the temperature of the medium. Varying the temperature of the medium may be carried out using a heater such as an on-chip heater. Varying the temperature of the medium may be carried out by environmental temperature. The temperature may be varied locally or globally. The method of the sixth aspect may include features corresponding to any features of the device of the first aspect, the method of the second aspect, the system of the third aspect, the computer vision product of the fourth aspect, and / or the method of the fifth aspect. Brief Description of the Drawings Certain embodiments of the present invention will now be described, by way of example, with reference to the accompanying drawings in which: Figure 1 is a schematic perspective view of a first optical image processing system comprising a device that comprises a medium; Figure 2A is a schematic plan view of a portion of the medium of Figure 1; Figure 2B is a schematic cross-section view along the line A-A' in Figure 2A. Figure 3 is a schematic plan view of a medium comprising spatially separated photonic networks; Figure 4 is a schematic side view of a medium disposed on a surface that is not flat; Figure 5 is a schematic side view of a medium that is a multi-layered heterostructure; Figure 6 is a schematic side view of a medium comprising voids; Figure 7 is a schematic side view of a medium comprising suspended structures; Figure 8 is a schematic perspective view of a medium that is 3D patterned; Figure 9 is a process flow diagram of a method of fabricating a device; Figure 10 is a schematic block diagram of a second optical image processing system; Figure 11 is a schematic block diagram of a third optical image processing system; Figure 12 is a schematic block diagram of a fourth optical image processing system; Figure 13 is a schematic side view of a device comprising electrodes for training; Figure 14 is a schematic side view of a device comprising nanomagnets for training; Figure 15 is a process flow diagram of a method of training a device; Figure 16 shows examples of correct (top) and incorrect (bottom) classifications; Figure 17 shows a test accuracy score of 97.7% determined using a device and by the MNIST handwritten digit classification task; Figure 18A shows an exemplary image (Figure 18A) to be usedin a comparison of optical physical computing using a device provided with raw image data with logistic regression performed on the same raw image data; Figure 18B is a photonic network to which the image of Figure 18A is provided; Figure 18C is a spatially-resolved spectrum from the photonic network of Figure 18B; Figure 18D shows a test accuracy score of 98.15% determined using spatially-resolved spectra including the spectrum of Figure 18C from a device; Figure 18E shows a test accuracy score of 90.5% determined by logistic regression on raw image data; Figure 18F is a perspective view illustration of optical physical computing; Figure 19 is a schematic perspective view of a device receiving spatially-structured illumination that is a spatially-structured response from another device; Figure 20 is a schematic block diagram of a fifth optical image processing system; Figure 21 is a schematic block diagram of a sixth optical image processing system; Figure 22 is a schematic block diagram of a seventh optical image processing system; Figure 23A Is experimental data showing an input image; Figures 23B and 23C are experimental data showing extracted feature / edge maps provided by two different optical modes; Figure 23D is experimental data showing feature / edge maps being extracted by multiple optical modes, where logistic regression has been used to train which combination of modes are used to extract image edges; and Figures 24A to 24D are simulated data illustrating detecting different edges (top / right / left / bottom) using four different optical modes, colour bars correspond to presence of the desired edge with 1 (white) indicating presence of the desired edge and 0 (black) indicating absence of the desired edge. Detailed Description of Certain Embodiments In the following, like parts are denoted by like reference numerals. Introduction Herein, devices for use in optical image processing are described. The devices can be driven by projecting spatially-structured illumination including input image data onto a medium of the device. As a result of the interactions of the spatially-structured illumination with and within the medium, a spatially-structured response is output. By modulating the interactions of the spatially-structured Illumination with and within the device, or alternatively by processing the input image data, the devices can be used to perform neuromorphic optical computing. First optical image processing system 1 Referring to Figure 1, a first optical image processing system 1 (hereinafter referred to as the "first system") is shown. The first system 1 includes a device 2 for optical image processing, a first light source 3 for providing spatially-structured illumination 4 including input image data to the device 2, and an imaging spectrometer 5 for detecting a spatially-structured response 6 from the device 2. The device 2 includes a medium 7 and a substrate 8. The medium 7 is disposed on a surface 9 of the substrate 8 and can support overlapping optical modes 10. In response to receiving the spatially-structured illumination 4, a set of overlapping optical modes 10 are excited in the medium 7 and output the spatially-structured response 6. The medium 7 includes a first photonic network 11 in which the input image data is processed and from which the spatially-structured response 6 is output. Referring also to Figures 2A and 2B, a portion of the first photonic network 11 is shown. The first photonic network 11 comprises a plurality of network links 12i, 12?, ..., 12n that are radiatively coupled at nodes 13i, 13?, ..., 13n and is disposed on the flat surface 9. The spatially-structured response 6 can include contributions arising from the spatially-structured illumination 4 interacting with the medium 7 in a variety of ways. First, the spatially-structured response 6 can includes a contribution arising from scattering of the spatially-structured illumination 4 from the medium 7. Secondly, the spatially-structured response 6 can include a contribution from photoluminescence, that is, absorption of the spatially-structured illumination 4 by the medium 7 leading to emission of photons by florescence or phosphorescence. Thirdly, the spatially-structured response 6 can include a contribution from stimulated emission processes such as lasing or amplified spontaneous emission ("ASE") that driven by the interaction of the spatially-structured illumination 4 with the medium 7. The set of overlapping optical modes 10 includes two or more optical modes that are lasing modes. The spatial-structuring of the spatially structured illumination 4 can unbalance mode completion between the lasing modes. This need not be the case, however, and the set of optical overlapping modes 10 may instead comprise only optical modes which are not lasing modes. Resonant modes capable of leading to lasing can be provided by using the first light source 3. The first light source 3 can include a suitable pulsed pump laser coupled into an optical microscope. The suitability of a pulsed pump laser is determined, at least in part, by whether it can provide a pulse fluence that is sufficiently large to induce lasing within the medium 7. In the case that the medium 7 includes a semiconductor material, the suitable pulsed pump laser can be a femtosecond pulsed laser. In the case that the medium 7 includes a polymer material, the suitable pulsed pump laser can be a nanosecond pulsed laser. By performing physical transforms in a parallel fashion and intrinsically using its internal (optical and optoelectronic) physics, the device 2 can have a higher processing speed than software networks. In particular, the speed of stimulated emission processes can mean that the device 2 can process an input image in as little as around 10 ps. Moreover, since the device 2 processes data taking the form of light, it can have a higher energy efficiency than conventional CMOS / GPU hardware. Spectra provided by the imaging spectrometer 5 based on the spatially-structured response 6 are sensitive to small variations in the spatially-structured illumination 4. Shifts in these spectra can form the basis on which image processing AI tasks can be performed using the device 2. The same input data, in the form of spatially-structured illumination 8, be provided to the device 2 multiple times, with transformations such as translations, rotations, reflections, spatial enlargement or shrinking. Providing the same input data to the device 2 multiple times with such transformations and combining the spatially-structured responses 6 for each time the input data is provided can help to enhance the computational performance (that is, the test accuracy score determined by the MNIST handwritten digit classification task) of the device 2. In modes of operation based on lasing, the power range for computing tasks is defined by measuring lasing light-in light-out curves for the device 2. The minimum power is defined by the lasing threshold (at which there is at least one lasing mode in the spectrum of the spatially-structured response 6) and the maximum power is defined by the power at which there is a maximum number of lasing modes in the spectrum of the spatially-structured response 6 or by the power at which there is a drop in light intensity output due to heating / material degradation, whichever of the two is lower. In these modes, the power range is typically well above the minimum power and well below the maximum power. The network links 12i, 122, ..., 12n are radiatively coupled by way of being in physical contact with each other at the nodes 13:, 132, ..., 13n. This need not be the case, however, and at least a sub-set of the network links 12i, 122, ..., 12n can be spatially separated from other network links at the nodes 13i, 132, ..., 13n by a distance less than a design wavelength at which the device 2 is configured to operate. In this way, direct physical contact between the network links 12i, 122, ..., 12n is not required. The network links 12i, 122, ..., 12n are active photonic waveguides. This need not be the case, however, and the network links 12i, 122, ..., 12Nmay include components other than the active photonic waveguides or may include low-dimensional features such as quantum wells instead of active photonic waveguides. Referring to Figure 3, a second example of the device 2 is shown. The medium 7 can include a second photonic network 112 that is spatially separated from the first photonic network 11 by a distance greater than the design wavelength. In this way, the device 2 can provide a set of parallel outputs. Combining parallel outputs can help to improve the computational performance of the device 2. Referring to Figure 4, a third example of the device 2 Is shown. The medium 7 can be disposed on a substrate 8 and on a surface 14 that is not flat. The surface 14 can have a wide variety of profiles that are not flat. In particular, the surface 14 may include flat portions connected by sloped portions or may be undulant. Furthermore, the surface 14 may not be a surface of the substrate 8 and may instead be defined by the combination of the substrate 8 and a patterned layer (not shown) such as a set of gate electrodes of a set of magnetic elements interposed between the substrate 8 and the medium 7 (best illustrated in Figure 13). Spatial structuring of the medium 7 Referring to Figure 5, a fourth example of the device 2 Is shown. The medium 7 can be a multi-layered heterostructure Including medium layers 15i, 152, ..., 15n. The medium layers 15i, 152, ..., 15n can have different thickness and / or composition and are not required to be planar. Referring to Figure 6, a fifth example of the device 2 is shown. Voids 16 may be arranged non-uniformly within the medium 7. The voids 16 may have a regular shape, such as an ellipsoidal shape, or an irregular shape. The voids 16 are not required to have the same shape and can have any shape. Referring to Figure 7, a sixth example of the device 2 is shown. Voids 16 can be arranged non-uniformly within the medium in such a way that suspended structures 17 are defined. The suspended structures 17 need not be regularly shaped. Referring to Figure 8, a seventh example of the device 2 is shown. The medium 7 need not be planar. Expressed differently, the medium 7 may have a 3D patterning in that it has both an in-plane texture and an out-of-plane texture. In particular, expanded polystyrene foam beads can be used to provide so-called "inverse opal" style 3D patterning. Alternatively, the 3D patterning can be provided via optical interference writing, or via nanoscale 3D printing such as so-called "two-photon lithography" or focused ion-beam deposition, or via 3D nanoscale self-assembly. Hereinbefore, it has been described that the medium 7 has a spatially non-uniform physical structure. This need not be the case, however, and the medium 7 can have a physical structure which is spatially-unform along three orthogonal directions by way of it having a complex refractive index which is spatially-non-uniform. In particular, the first photonic network 11 can replaced with a uniform thin film having a complex refractive index and a second light source (not shown) can be used to provide the spatially non-uniform complex refractive index by projecting a spatially-structured light beam onto the uniform thin film. Expressed differently, the first light source 3 is used to provide the spatially-structured illumination 4 as a probe and the second light source can be is used to provide the spatial structuring of the complex refractive index of the medium 7. Device fabrication The medium 7 can include a wide variety of semiconductor materials. In particular, the medium 7 can be include a III-V semiconductor material such as indium phosphide (InP) or an organic semiconductor material such as Rhodamine 6G dye. Thus, the medium can be a solid-state gain medium or a liquid-phase dye medium. The device 2 can be fabricated in a wide variety of ways, such as by lithography on a thin III-V semiconductor layer that is bonded on a SiOz-coated Si substrate, by electrospinning dye-mixed polymer nanofibers onto a TEM grid and post-annealing to fuse the network links 12i, 122, ..., 12n together to form at least one photonic network 11, 112, or by nanoimprint lithography of soft polymers doped with laser dye. Media such as the medium 7 can be cheap to produce, near-infinitely configurable, and readily compatible with widespread semiconductor and polymer fabrication techniques. Referring to Figure 9, a method of the device 2 will now be described. The method includes providing a first substrate having a principal surface (step Sl.l), providing a hetero-epitaxial wafer having an intermediate sacrificial layer interposed between a second substrate and an upper layer (step SI.2), bonding the upper layer to the principal surface of the first substrate (step SI.3), and etching the sacrificial layer to remove the second substrate and the sacrificial layer to form a bonded medium that includes the first substrate and the upper layer by (step SI.4). The method may then include lithographically patterning structures onto the surface of the bonded medium from which the sacrificial layer was etched (step SI.5). The bonded medium is an example of the device 2. In particular, the first substrate can be the substrate 8 and the principal surface can be the surface 9 or the surface 14. The first substrate may include or be a silicon (Si) wafer. The principal surface may be a surface of the silicon wafer or a surface of a silicon dioxide (SiCh) layer deposited on the silicon. The second substrate may include an indium phosphide (InP) wafer. The intermediate sacrificial layer may include indium gallium arsenide (InGaAs). The upper layer may include indium phosphide (InP). Second optical image processing system 18 Referring to Figure 10, a second optical image processing system 18 (hereinafter referred to as the "second system") Is shown. The second system 18 is different to the first system 1 in that it uses an external source of spatially structured illumination 19 instead of in place of the first light source 3 and includes a spectrometer 20 that is not required to be an imaging spectrometer such as the imaging spectrometer 5. Expressed differently, the spectrometer 20 Is not required to be spatially-resolved. In the second system 18, the device 2 receives spatially-structured illumination 4 from the source of spatially structured Illumination 19 and outputs Its spatially-structured response 6 to the spectrometer 20. The readout of the spectrometer 20 is provided to a computer system 1 for processing. Because the spectrometer 20 is not required to be spatially-resolved, it can provide a readout which corresponds to the spatially-structured response 6 from the whole of the medium 7. In the case that the spectrometer 20 is an imaging spectrometer such as the imaging spectrometer 5, however, the spectrometer 20 can provide a readout which corresponds to the spatially-structured response 6 from at least one defined spatial sub-region of the medium 7. In the spectrometer 20, the spatially-structured response 6 can be collected by a microscope objective (not shown), filtered to remove a contribution corresponding to the spectrally-structured illumination 4, spectrally dispersed using a grating (not shown), and focussed into a CCD camera (not shown). Spatially-resolved readout can be enabled by including an arrangement of lenses (not shown) between the medium 7 and the spectrometer 20. The second system 18 can use a neutral density filter (not shown) to maintain a pump power of the spatially-structured illumination 4 by monitoring and correcting the pump power with respect to reference power values. Implementing such real-time correction can help to improve computational performance. Third optical image processing system 22 - free-space implementation Referring to Figure 11, a third optical image processing system 22 (hereinafter referred to as the "third system") is shown. The third system 22 has a free-space implementation in which light is projected through free space between its component parts. The third system 22 is different to the second system 18 in that includes a digital micromirror device ("DMD") 23 arranged in a reflection geometry. In particular, the DMD 23 Is arranged to receive input light 24 from a source of illumination 25 external to the third system 22, and to provide the spatially-structured illumination 4 to the device 2 via a beam splitter 26 and an objective 27. The spatially-structured response 6 of the device 2 is detected by the spectrometer 20, the readout of which is provided to the computer system 21 for processing. Using a DMD to impart a spatial structure such as a spatial structure corresponding to an image onto input light such as the input light 24 can help to make the third system 22 cheaper and easier to manufacture. This is because DMD components are widely available and industrially produced. Using a DMD 23 can also help to make the third system 22 suitable for operating at GHz and even THz frequencies. Fourth optical image processing system 28 - on-chip implementation Referring to Figure 12, a fourth optical image processing system 28 (hereinafter referred to as the "fourth system") is shown. The fourth system 28 is different from the third system 22 in that it has an on-chip implementation rather than a free-space implementation. The fourth system 28 is also different from the third system 22 in that it uses a spatial light modulator ("SLM") 29 in a transmission geometry instead of a DMD such as the DMD 23 to provide the spatially-structured illumination 4. The component parts of the fourth system 28, namely the SLM 29, the device 2, and the spectrometer 20, are arranged in an on-chip package (herein also referred to as a "photonics chip"). In the on-chip package, the component parts of the fourth system 28 are arranged in a multi-layer vertical stack on a flat substrate such as a silicon (Si) chip. The on-chip package being flat can help to make the system 28 easier to integrate integration into larger assemblies. The SLM 29 is used to impart a spatial structure on the input light 24 provided by the source of illumination 25. The source of illumination 25 can be a spatially-uniform flat form factor light source such as a microstrip laser. The SLM need not be used to impart the spatial structure to the spatially-structured Illumination 4, however, and the spatially-structured illumination 4 can instead be provided directly from a source of illumination 25 such as a nanolaser array, a vertical-cavity surface-emitting laser ("VCSEL") array, or a miniature light-emitting diode ("LED") array, that is, a source of illumination 25 capable of directly outputting the spatially structured response 4. Neuromorphic computing schemes The device 2 described herein can form the basis for implementing one or more neuromorphic computing schemes. Reservoir computing The device 2 can be used as a reservoir for reservoir computing. In this scheme, the device 2 is not required to be trainable and the input data in the spatially-structured illumination 4 translates into a spatiotemporal gain profile In the device 2 due to absorption of the spatially-structured illumination 4 by the medium 7. Light emitted from the medium 7 is guided in the medium 7 and propagates across multiple paths within the medium 7. Constructive interference of light across the at least two closed paths within the medium can result in the optical modes 10, which can be amplified with optical gain. Lasing from these modes 10 occurs at the threshold when optical gain for a mode balances the mode losses. The optical modes 10 have a non-zero spatial overlap and so can compete for gain. When two or more optical modes 10 lase, mode competition due to spatial hole burning and non-linear coupling can result in non-linear variation of the mode intensity and the wavelength with the pump-power of the spatially-structured illumination 4 across the medium 7. In this way, a system including the device 2 may perform multiple linear non-linear operations on the input data and output unique spectra such as lasing spectra formed from these operations. Reservoir computing can be performed by processing these spectra, that Is, the output of the reservoir. In particular, the spectral output can be processed by regression to directly produce a computational output. The reservoir computing scheme can be trained for a given computational task by optimising the processing of these spectra for one or more given inputs. Input data can pre-processed to take the form of the spatially-structured illumination 4. In particular, analogue time-series data can be converted to grayscale pixel illumination values and illuminated onto the medium 7. One way to pre-process the input data is to pre-process the input data as a matrix of positive values, binarize the pre-processed input data into pattern projections on a DMD such as the DMD 23 or a SLM such as the SLM 29, and to project the spatially-structured illumination 6 from the DMD or SLM onto the device 2. Input data pre-processing can include pixel dithering and / or noise representation to enable grayscale and / or analogue value input. These types of input data preprocessing can help to improve computational performance. Deep neural networks The device 2 can be used in a fully trainable deep neural network ("DNN") for inference. In particular, the device 2 can be configured as a layer or layers in a neural network such as a convolutional neural network ("CNN"). In this scheme, the device 2 is required to be trainable. By projecting image data such as image fragments onto the medium 7, the device 2 can be used as a convolutional kernel. In this scheme, the spectral output from the spectrometer 20 is passed to another network layer for further processing instead of being processed by regression to directly produce a computational output. The spatially-resolved response from a first network layer including the device 2 can be projected as an image onto a further network layer (best Illustrated in Figure 19). The further network layer can be a distinct physical network with different properties to the first network layer. The further network layer can include or be hardware such as the device 2, a memristive network, or a nanomagnetic network. The first network layer may have software and / or hardware layers before it, and instead of image data, the input data may be the output of preceding or subsequent network layers (best illustrated in Figures 21 and 22). In this way, the device 2 can be used as a pre / post-processor as part of a larger computing system. This can add memory and thereby can allow the device to be used for video processing. Output data can be post-processed In external circuits or an external CPU such as the computer system 21. Output data can be extracted directly from the spatially-structured response 6, for example using spatially-resolved spectra such as optical lasing spectra read out from the spectrometer 20. Training The device 2 can be trained in wide a variety of ways. A first mode for training the device 2 uses the second light source (not shown) to provide spatial structuring of the complex refractive index of the medium 7. Referring to Figure 13, a device 2 including electrodes 30 for training the medium 7 interposed between the medium 7 and the substrate 8 is shown. The device 2 need not include this specific arrangement of electrodes 30, however, and the medium 7 may instead be Interposed between the electrodes 30 and the substrate 8. Expressed differently, the electrodes 30 may be disposed on an upper surface of the medium 7 opposite to the substrate 8. Alternatively, the electrodes 30 may be arranged to be adjacent to the medium 7, that is, offset from the medium 7 in a plane in which the medium 7 at least predominantly Iles. The electrodes 30 can be arranged to define a 2D grid (or "pixel grid"). This arrangement can improve the ease with which the electrodes 30 can be addressed. In a second mode for training the device 2, the electrodes 30 can be used to apply a locally reconfigurable electric field to the medium 7 and thereby can allow for local electrical control of the medium 7. Referring to Figure 14, a device 2 including magnetic elements 31 disposed on an upper surface of the medium 7 is shown. The device 2 need not include this specific arrangement of magnetic elements 31, however, and the medium 7 may instead be interposed between the magnetic elements 31 and the substrate 8. Expressed differently, the magnetic elements 31 may be disposed on an upper surface of the medium 7 opposite to the substrate 8. Alternatively, the magnetic elements 31 may be arranged to be adjacent to the medium 7, that is, offset from the medium 7 in a plane in which the medium 7 at least predominantly lies. The magnetic elements 31 can be arranged to define a 2D grid. This arrangement can improve the ease with which the magnetic elements 31 can be addressed. In a third mode for training the device, the magnetic elements 31 can be used to apply a locally reconfigurable magnetic field across the medium 7 and thereby can allow for local magnetic and / or magneto-plasmonic control of the medium 7. As will be discussed hereinafter, a fourth mode for training the device 2 includes applying heat to the medium 7. Referring to Figure 15, a method of training the device 2 will now be described. The method includes: initiating training (step S2.1); determining a loss function of the device 2 (step S2.2); making a determination of whether the loss function has reached a stable minimum (step S2.3); in the case of a negative determination, modulating an optical response in the device 2 (step S2.4) and repeating steps S2.2 and S2.3 in that order; in the case of a positive determination, ending training (step S2.5). Determining a loss function of the device 2 can include assessing task performance of the device for a given task and generating the loss function based on the task performance. Modulating an optical response in the device 2 means that the interaction(s) of the spatially-structured illumination 4 with and within the medium 7 are modified, resulting In a change in spatially-structured response 6 provided for a given spatially-structured illumination 4 input. In the case that accuracy is low, the loss function is high. In the case that accuracy is high, the loss function is low. The higher the loss, the greater the change in the physical response between iterations of the training. After a number of iterations of the method have been carried out, the loss reaches a stable minimum and the computational performance of the device 2 can be described as optimised for the given computational task. In other words, in the case that the loss is at a stable minimum, the device 2 can be described as being trained for the given computational task. Modulating the optical response in the device 2 includes adjusting a field applied to the medium 7. The field can be applied to the medium 7 locally or globally. Expressed differently, the field can be applied to only a portion of the medium 7 or to the entirety of the medium 7. In the first mode for training the device 2, the field takes the form of a mask taking the form of the spatially-structured light beam from the second light source that is projected onto at least a part of the medium 7. The spatially structured light beam from the second light source can be spatially-structured by a DMD or by a SLM, by optical interference or holography, or by spatial focussing. In the first mode for training the device, a trained grid of pixels in the mask are superimposed over the spatially-structured illumination 4 on the medium 7 to modulate the complex refractive index of the medium 7. Expressed differently, in the first mode of training the device 2, a grid of pixels in a mask taking the form of the spatially-structured light beam from the second light source is used to train the medium 7. In the second mode for training the device 2, the electrodes 30 are used to apply a field that is an electrical field to the medium 7. In modes of operation of the device 2 which include lasing modes, the electrodes 30 can be used to apply a reconfigurable electric field that modulates a lasing threshold of the medium 7. In modes of operation of the device 2 based on photoluminescence, the electrodes 30 can be used to apply a reconfigurable electric field that modulates how the spatially-structured illumination 4 interacts with different regions of the medium 7. In this way, different regions of the medium 7 can respond with higher or lower, and / or shifted, sensitivity to the spatially-structured Illumination 4. In the third mode for training the device 2, the field is a magnetic field and the magnetic elements 31 are used to exert magnetic and / or magnetoplasmonic control 7 over the medium 7. Each magnetic element 31 has a magnetic state that locally modifies how much of the spatially-structured illumination 4 is absorbed and / or enhanced. The magnetic states can be reconfigurably programmed in a wide variety of ways such as by an applied magnetic field, by current / voltage control, and / or by all-optical magnetic switching. In the fourth mode for training the device 2, the field takes the form of heat applied to the medium 7. The heat can be applied to the medium 7 using a heater such as an on-chip heater or by local heating such as by laser heating. Alternatively, the heating may be provided by environmental temperature in such a way that the device 2 is a sensor that varies its physical dynamics in response to changes in its environment. Such a sensor can be applied in atmospheric or weather sensing and / or in vehicular or process control sensing. Computational performance of the device 2 The computational performance of the device 2 has been assessed using the MNIST handwritten digit classification task. Referring to Figure 16, examples of correct (top) and incorrect (bottom) classifications of handwritten digits are shown. The errors are dominated by ambiguous digits that human readers may struggle with. Referring also to Figure 17, a test accuracy score of 97.7% determined using the device 2 and by the MNIST handwritten digit classification task is shown. The test accuracy score of 97.7% was calculated from the confusion matrix shown in Figure 17 and based on classifying 10,000 images, 233 of which were classified incorrectly and 9767 of which were classified correctly. Due to floating point precision, the sum of the values in each row and column shown is not always exactly 1.000. For example, the sum of the values in the row having an actual label of 0 is equal to 1.000, the sum of the values in the row having an actual label of 1 is 1.001, and the sum of the values in the row having an actual label of 8 is 0.999. Referring also to Figures 18A to 18F, a comparison of optical physical computing using a device provided with raw image data with logistic regression performed on the same raw image data is shown. In Figure 18A, an exemplary image used in the MNIST handwritten digit classification task is shown. In Figure 18B, a plan view of an example of the photonic network 11 is shown. The image shown in panel a was one of the 10,000 images projected onto the device 2 whilst the MNIST handwritten digit classification task was being performed. In Figure 18C, a spatially-resolved spectrum read out by the imaging spectrometer 5 is shown. The upper line profile indicates data from a defined spatial sub-region of the medium 7 and corresponds to the region of Figure 18C within the light rectangular box The lower line profile indicates data from the entirety of the medium 7 and corresponds to the region of Figure 18C within the dark square box. Logistic regression is performed on the readout from the imaging spectrometer 5 by the computer system 21 and, in Figure 18D, an accuracy of 98.15% is calculated from the confusion matrix shown. In panel e, logistic regression on the same raw image data as was used in the MNIST handwritten digit classification task was performed directly and a lower accuracy of 90.5% is calculated from the confusion matrix shown. In contrast to the values in the confusion matrix of Figure 18D, the values in the confusion matrix shown of Figure 18E are not normalised. Thus, the device 2 can provide a test accuracy score, measured by the MNIST handwritten digit classification task, equal to 98.15%. This test accuracy score is an improvement on the test accuracy score that can be achieved by directly performing logistic regression on the same raw image data. Specific examples of neuromorphic computing schemes including the device 2 DNN Hereinbefore it has been described that device 2 can be used in a fully trainable DNN for inference. In particular, the device 2 can be configured as a layer or layers in a neural network such as a CNN. Referring to Figure 19, a first network layer li and a second network layer 12 are shown. The first network layer li and the second network layer 12 are each modified versions of the first system 1 and each Include a device 2i, 22. The first network layer li is different to the first system 1 only in that it does not include an imaging spectrometer 11. The second network layer I2 is different to the first system 1 in that the spatially structured illumination provided to the device 22 is the spatially-structured response 61 of the first network layer li. The spatially structured response 61 of the first network layer li and the spatially structured response 62 of the second network layer I2 are both detected by the imaging spectrometer (not shown) of the second network layer I2. The first network layer li and the second network layer I2 are connected in series. This need not be the case, however, and further network layers (not shown) can be included, and network layers can be connected in a feedback loop with the spatially-structured response 62 of the second network layer used as the spatially-structured illumination 4i of the first network layer li. Network layers can be connected to form feed-forward neural networks such as CNNs or can be connected to form recurrent neural networks. It need not be the case that the first network layer li and the second network layer I2 include separate devices 2i, 22. Instead, the network layer li can include the first photonic network 11 and the second network layer I2 can include the second photonic network II2. The first photonic network 11 and the second photonic network II2 are included In the same device 2 (best Illustrated in Figure 3). Combining parallel outputs, for example from the first network layer li and the second network layer I2, can help to improve computational performance. Fifth optical image processing system 32 - stand-alone processor Referring to Figure 20, a fifth optical image processing system 32 (hereinafter referred to as the "fifth system") is shown. The fifth system 32 comprises only one device 2. Thus, the fifth system 32 is implemented as a stand-alone processor. The fifth system 32 can be used is an example of the first system 1 which is implemented as a stand-alone processor and comprises only one device 2. The device 2 of the fifth system 32 can be used either as a reservoir or as a network layer in a DNN. Sixth optical image processing system 33 - pre-processor Referring to Figure 21, a sixth optical image processing system 33 (hereinafter referred to as the "sixth system") is shown. The sixth system 33 is a generalised version of the example DNN example described hereinbefore as including the first network layer li and the second network layer 12. In particular, the sixth system 33 includes two network layers connected in series, only the first of which includes a device 2. In the sixth system 33, the device 2 receives the spatially-structured illumination 4 as an input and provides the spatially-structured response 6 as an output. The second network layer 34 receives the spatially-structured response as an input and provides an output 35. In this way, the device 2 is used within the system 33 as a pre-processor. Seventh optical image processing system 36 - post-processor Referring to Figure 22, a seventh optical image processing system 36 (hereinafter referred to as the "seventh system") is shown. The seventh system 36 is a modified version of the sixth system 33 in which the device 2 is used as a post-processor instead of a pre-processor. In the seventh system 36, a first network layer 37 receives an input 38 and provides an output 39. The device 2 receives the spatially-structured illumination 4 as an input, the spatially-structured illumination provided based on the output 39 and provides the spatially-structured response 6 as an output. In these ways, the device 2 can be used as a pre / post-processor as part of a larger computing system. As described hereinbefore, using the device 2 as a network layer in combination with a further network layer including for example a memristive network or a nanomagnetic network can add memory and thereby can allow the device 2 to be used for video processing. In the sixth system 33 and the seventh system 36, network layers need not be connected in series, and two or more network layers may be connected in parallel. Edge / feature detection The device 2 can be used to perform convolutional edge detection directly using its internal physics, with no requirement for regression or post-processing. In particular, the first system 1 can serve as both the convolutional kernel layer and classifier layers of a convolutional neural network. The kernel layer can be used as a standalone edge-detection / feature-extraction system, which treats the optical power of each lasing mode as a separate feature extraction kernel. A first optical mode at one wavelength may respond more strongly when vertical edges are present, a second optical mode at another wavelength may respond more strongly when horizontal edges are present. In this way, mode powers can be used to construct feature maps, either with or without prior training / regression of the kernels. The edge / feature detection functionality can act as a layer within a larger image processing and / or computing system, such as an initial kernel layer of a convolutional neural network. Alternatively, the edge / feature detection functionality can act as a standalone processor which extracts edges / features from images and returns maps of the edges / features as its output. Referring to Figures 23A to 23C, examples of an input image (the letters ICL) (Figure 23A) and extracted feature / edge maps provided by two different optical modes 10 (Figures 23B and 23C) are shown. Each of Figures 23B and 23C show contrast corresponding to edges or features of the input image of Figure 23A. Further experimental data of other examples of edge-detection in which regression has been used to train which of the optical modes are used to train edge maps is shown in Figure 23D. A single device 2 can provide many kernels in parallel by multiplexing kernels In the spectral frequency space via the mode / kernel correspondence. Referring also to Figures 24A to 24D, examples of detecting all four edges using four different optical modes 10 (one In each of Figures 24A to 24B) is shown. Each of Figures 24A to 24D show contrast indicating at least one of the edges of a simulated square. Thus, the device 2 can be for edge detection and feature detection. Applications The device 2 can be used in a wide variety of computer vision applications ranging from autonomous piloting of vehicles to object detection. In a first example, the device 2 is used in a self-driving car in which image data from car mounted cameras is projected optically onto at least one device 2 that is preconfigured and trained for hazard detection. In this application, where response time is critical, the rapid processing speed of the device 2, relative to conventional software neural network processing, can be particularly beneficial. A self-driving car including the device 2 can respond to observed oncoming hazards by providing a spectral readout which includes one or more characteristic features indicative of a warning, that spectral readout acting as a trigger for activating evasive manoeuvres in the selfdriving circuits, keeping passengers and pedestrians safe. In related examples, the device 2 can be used in another other autonomous vehicle, such as an autonomous drone, an autonomous aircraft, or an autonomous sea vessel. By providing a processor such as the computing system 21 for determining variations in the spectral readout of the spectrometer 20, the device 2 can be particularly suitable for use in such remote non-data-centre use cases. In a second example, the device 2 is included in a CCTV camera processing system which can respond to observing certain objects, such as crowds or faces, by providing a spectral readout which includes one or more characteristic features indicative of the object(s) observed, similarly to how the warning spectra is provided in the self-driving car. In related examples, the device 2 can be included in another processing system, such as an industrial manufacture control system, a medical imaging system, or a surgical control system for example for processing endoscope camera image, and / or video data. The device 2 can also be used in a wide variety of other machine learning applications in which data for non-linear processing can be mapped onto an image. Modifications It will be appreciated that various modifications may be made to the embodiments hereinbefore described. Such modifications may involve equivalent and other features which are already known in the fields of optical image processing and neuromorphic computing hardware, and which may be used instead of or in addition to features already described herein. Features of one embodiment may be replaced or supplemented by features of another embodiment. 5 Although claims have been formulated in this application to particular combinations of features, it should be understood that the scope of the disclosure of the present invention also includes any novel features or any novel combination of features disclosed herein either explicitly or implicitly or any generalization thereof, whether or not it relates to the same invention as presently claimed in any claim and whether or 10 not it mitigates any or all of the same technical problems as does the present invention. The applicants hereby give notice that new claims may be formulated to such features and / or combinations of such features during the prosecution of the present application or of any further application derived therefrom.

Claims

1. A device for use in optical image processing, the device comprising:• a medium capable of supporting overlapping optical modes such that, in response to spatially-structured illumination, a set of overlapping optical modes in the medium are non-uniformly excited and output a spatially-structured response.

2. The device of claim 1, wherein the medium comprises:• an optical gain medium.

3. The device of claim 1 or 2, wherein the medium is spatially non-uniform.

4. The device of any one of claims 1 to 3, wherein the medium is configured as apanel.

5. The device of any one of claims 1 to 4, wherein the medium comprises a first photonic network.

6. The device of claim 5, wherein the medium comprises a second photonic network that is spatially separated from the first photonic network.

7. The device of any one of claims 1 to 6, wherein the medium comprises photonic waveguides.

8. The device of any one of claims 1 to 7, wherein the medium comprises lowdimensional structures.

9. The device of claim 8, wherein the low-dimensional structures comprise quantum wells.

10. The device of any one of claims 1 to 9, wherein the medium comprises a semiconductor material.

11. The device of claim 10, wherein the semiconductor material is an inorganic semiconductor material.

12. The device of claim 10 or 11, wherein the semiconductor material is a III-V semiconductor material.

13. The device of claim 10, wherein the semiconductor material is an organic semiconductor material.

14. An optical image processing system, comprising: ■ the device of any one of claims 1 to 13; and • an imaging spectrometer configured to detect the spatially-structured response.

15. The optical image processing system of claim 14, wherein the imaging spectrometer is configured to perform readout on defined spatial sub-regions of the medium.

16. The optical image processing system of claims 14 or 15, further comprising: ■ a neutral density filter configured to maintain a pump power of the spatially-structured illumination.

17. The optical image processing system of any one of claims 14 to 16, further comprising:■ a digital micromirror device configured to provide the spatially-structured illumination.

18. The optical image processing system of any one of claims 14 to 16, further comprising:■ a spatial light modulator configured to provide the spatially-structured illumination.

19. The optical image processing system of any one of claims 14 to 18, wherein the optical image processing system is configured for use as a non-trainable reservoir for data processing.

20. The optical image processing system of any one of claims 14 to 19, wherein the optical image processing system is configured for use as a fully trainable deep-neural network for inference.

21. A computer vision product, comprising:• the optical image processing system of any one of claims 14 to 20; and■ a processor configured to determine variations in spatially-resolved spectra provided by the spatially-resolved spectrometer.

22. The computer vision product of claim 21, wherein the computer vision product is a vehicle capable of being autonomously piloted or a system for object detection and classification.

23. A method of operating the device of any one of claims 1 to 13 or the optical image processing system of any one of claims 14 to 20, the method comprising: providing the spatially-structured illumination to the device; and detecting the spatially-structured response.

24. A method of training the device of any one of claims 1 to 13 or the optical image processing system of any one of claims 14 to 20, the method comprising: determining a loss function of the device, determining whether the loss function has reached a stable minima, and modulating an optical response in the device.

25. The method of claim 24, wherein modulating an optical response in the device comprises adjusting a field applied to the medium.36

Images