Neuromorphic circuit for physically producing a neural network and associated production and inference method

Abstract

The invention relates to a neuromorphic circuit (10) for physically producing a neural network. The neuromorphic circuit (10) comprises: - at least one component (26) that is capable of being excited according to a plurality of excitation eigenmodes with a respective population and of having a plurality of excitation configurations, - a unit (28) for configuring the component (26) capable of exciting the component (26) in order to obtain an excitation configuration chosen on the basis of a desired neural network architecture, and - a unit (30) for interrogating the component (26) capable of selectively modifying the populations of first excitation eigenmodes and of measuring the populations of second excitation eigenmodes, the first and second excitation eigenmodes being the excitation modes of the input neurons and the output neurons of the desired architecture, respectively.

Description

[0001] Neuromorphic circuit physically realizing a neural network and associated realization and inference method

[0002] FIELD OF THE INVENTION

[0003] The present invention relates to a neuromorphic circuit physically implementing a neural network. It also relates to an associated implementation and inference method.

[0004] BACKGROUND OF THE INVENTION

[0005] The development of the internet and connected sensors has led to the acquisition of considerable quantities of data. This phenomenon, often referred to as "big data," involves the use of computers to exploit all the data obtained. Such exploitation can be used in multiple fields, including automatic data processing, diagnostic assistance, predictive analysis, autonomous vehicles, bioinformatics, and surveillance.

[0006] To implement such exploitation, it is known to use machine learning algorithms that are part of programs that can be executed on processors such as CPUs or GPUs. A CPU is a processor, the acronym CPU coming from the English term "Central Processing Unit" while a GPU is a graphics processor, the acronym GPU coming from the English term "Graphics Processing Unit" literally meaning graphics processing unit.

[0007] Among the techniques for implementing learning, the use of formal neural networks, and in particular deep neural networks, is increasingly widespread, these structures being considered very promising due to their performance for numerous tasks such as automatic data classification, pattern recognition, automatic language translation and understanding, robotic control, automatic navigation, recommendation systems, anomaly detection, fraud detection, the study of DNA or the discovery of new molecules.

[0008] A neural network is generally composed of a succession of layers of neurons, each of which takes its inputs from the outputs of the previous layer. More precisely, each layer includes neurons taking their inputs from the outputs of the neurons in the previous layer. Each layer is connected by a plurality of synapses. A synaptic weight is associated with each synapse. It is a real number, which takes both positive and negative values. For each layer, the input of a neuron is the weighted sum of the outputs of the neurons in the previous layer, the weighting being done by the synaptic weights.

[0009] For an implementation in a CPU or a GPU, a Von Neumann bottleneck problem arises because the implementation of a deep neural network (with more than three layers and up to several dozen) involves using both the memory(s) and the processor while these latter elements are spatially separated. This results in a congestion of the communication bus between the memory(s) and the processor both while the neural network, once trained, is used to perform a task, and, even more so, while the neural network is being trained, that is, while its synaptic weights are being adjusted to solve the task in question with maximum performance.

[0010] It is therefore desirable to develop dedicated hardware architectures, combining memory and computing, to create fast, low-power neural networks capable of learning in real time.

[0011] It is known to produce neural networks based on CMOS technology. It is understood by the acronym "CMOS", complementary metal-oxide-semiconductor (acronym coming from the English expression "Complementary Metal-Oxide-Semiconductor"). The acronym CMOS refers to both a manufacturing process and a component obtained by such a manufacturing process.

[0012] A neural network based on optical technologies is also known.

[0013] Specifically, three architecture proposals are the subject of specific studies: CMOS neural networks and CMOS synapses, optical neural networks and optical synapses, and CMOS neural networks and memristive synapses. Memristive synapses are synapses using memristors. In electronics, a memristor (or memristance) is a passive electronic component. The name is a portmanteau of the two English words memory and resistor. A memristor is a non-volatile memory component, the value of its electrical resistance changing with the application of a voltage over a certain period of time and remaining at that value in the absence of voltage.

[0014] However, in each of these technologies, each neuron is several tens of micrometers wide. For CMOS and optical technologies, each synapse is also several tens of micrometers wide. As a result, on a limited surface area, such as an electronic chip, the number of neurons and synapses that can be integrated is limited, resulting in a reduction in the performance of the neural network.

[0015] For spiking neural networks, in addition to the aforementioned problems, there is also the difficulty of implementing training since the gradient backpropagation technique, which is currently the most effective, cannot be used directly.

[0016] To train spiking neural networks, it is known to use a principle derived from Hebb's rule according to which two neurons have a stronger synaptic connection when they emit impulses simultaneously. For example, a constraint is applied to each synapse so that the associated weight is increased if the two neurons connected to the synapse have each emitted an impulse with a time interval lower than a predefined threshold or is decreased otherwise.

[0017] However, because this technique does not solve the optimization of a global objective function, its accuracy is not always satisfactory.

[0018] SUMMARY OF THE INVENTION

[0019] There is therefore a need for a neuromorphic circuit capable of producing a large number of synapses and neurons and with reduced power consumption.

[0020] For this purpose, the description describes a neuromorphic circuit physically realizing a neural network, the neural network comprising a set of neurons connected by synapses, the neuromorphic circuit comprising:

[0021] - at least one component capable of being excited according to a plurality of specific excitation modes with a respective population, the component thus being capable of presenting several excitation configurations, each excitation configuration being defined by the specific excitation modes excited and their respective population, each pair of excitation modes being coupled by a respective coupling, each specific excitation mode being a neuron of the neural network and each coupling being a synapse of the neural network,

[0022] - a configuration unit for configuring the at least one component, the configuration unit being capable of exciting the component to obtain an excitation configuration chosen according to a desired neural network architecture, the desired architecture comprising input neurons and output neurons, and

[0023] - an interrogation unit of the at least one component, the interrogation unit of the component being capable of selectively modifying the populations of first specific excitation modes and of measuring the populations of second specific excitation modes, the first specific excitation modes being the excitation modes of the input neurons and the second specific excitation modes being the excitation modes of the output neurons.

[0024] According to particular embodiments, the neuromorphic circuit has one or more of the following characteristics, taken in isolation or in all technically possible combinations:

[0025] - the component is capable of being excited in a regime in which the amplitude of a coupling between each pair of eigenexcitation modes in each excitation configuration depends on the respective populations of each of the two eigenexcitation modes;

[0026] - the at least one component has two regimes, a first regime in which the amplitude of the coupling between each pair of natural excitation modes is independent of the population of each of the two natural excitation modes and a second regime in which the amplitude of the coupling between each pair of natural excitation modes depends on the population of each of the two natural excitation modes, the configuration unit being capable of exciting the component in the second regime;

[0027] - the configuration unit is capable of exciting the specific excitation modes according to an initial configuration different from the chosen excitation configuration, the component relaxing towards the chosen excitation configuration;

[0028] - the at least one component is a layer of ferromagnetic element;

[0029] - the ferromagnetic element is an iron and yttrium garnet;

[0030] - the at least one component is chosen from the list consisting of: a magnetic microstructure, a cavity, a metamaterial, and a superconducting resonator;

[0031] - the configuration unit is chosen from an optical excitation unit and an electrical excitation unit.

[0032] - the interrogation unit is chosen from: an optical excitation unit, and an electrical excitation unit.

[0033] The description also describes a method of physically realizing a neural network having a desired architecture, the neural network comprising a set of neurons connected by synapses, the method of physical realization comprising:

[0034] - a step of configuring at least one component of a neuromorphic circuit physically realizing a neural network having been configured, the at least one component being capable of being excited according to a plurality of specific excitation modes with a respective population, the component thus being capable of presenting several excitation configurations, each excitation configuration being defined by the excited specific excitation modes and their respective population, each pair of excitation modes being coupled by a respective coupling, each specific excitation mode being a neuron of the neural network and each coupling being a synapse of the neural network, the configuration step being implemented by a configuration unit of the at least one component,the configuration unit being part of the neuromorphic circuit and comprising the excitation of the at least one component to obtain an excitation configuration chosen according to the desired architecture.,

[0035] The description also relates to a method for inferring a neural network having a desired architecture, the neural network comprising a set of neurons connected by synapses, at least one component of a neuromorphic circuit physically realizing a neural network having been configured, the at least one component being able to be excited according to a plurality of specific excitation modes with a respective population, the component thus being able to have several excitation configurations, each excitation configuration being defined by the excited specific excitation modes and their respective population, each pair of excitation modes being coupled by a respective coupling, each specific excitation mode being a neuron of the neural network and each coupling being a synapse of the neural network, the configuration having been implemented by a configuration unit of the at least one component,the configuration unit being part of the neuromorphic circuit and comprising the excitation of the at least one component to obtain an excitation configuration chosen according to the desired architecture, the inference method comprising:,

[0036] - a step of selective modification of the populations of first proper excitation modes, the first proper excitation modes being the excitation modes of the input neurons, the modification step being implemented by an interrogation unit of the neuromorphic circuit, and

[0037] - a step of measuring the populations of second proper excitation modes, the second proper excitation modes being the excitation modes of the output neurons, the measurement step being carried out after reaching a stable state for the at least one component, the measurement step being implemented by the interrogation unit.

[0038] In this description, the expression "suitable for" means indifferently

[0039] “adapted for”, “suitable for” or “configured for”. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Characteristics and advantages of the invention will appear on reading the description which follows, given solely by way of non-limiting example, and made with reference to the appended drawings, in which:

[0041] - Figure 1 is a schematic representation of a neuromorphic circuit,

[0042] - Figure 2 is a schematic representation of an example of a neural network,

[0043] - Figure 3 is a flowchart of an example implementation of an inference method using the neuromorphic circuit of Figure 1, and

[0044] - Figure 4 is an example of an experimental realization of the neuromorphic circuit of Figure 1.

[0045] DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

[0046] GENERAL NEUROMORPHIC CIRCUIT

[0047] Referring to Figure 1, a neuromorphic circuit 10 is described.

[0048] As its name suggests, the neuromorphic circuit 10 is a circuit physically realizing a neural network 12.

[0049] In other words, the neuromorphic circuit 10 is a physical circuit adapted to carry out the operations of a neural network 12.

[0050] In particular, the neuromorphic circuit 10 is capable of implementing a neural network 12 as shown diagrammatically in FIG. 2.

[0051] The neural network 12 described is a network comprising an ordered succession of layers 14 of neurons 16, each of which takes its inputs from the outputs of the previous layer 14.

[0052] By definition, in biology, a neuron, or nerve cell, is an excitable cell constituting the basic functional unit of the nervous system. Neurons ensure the transmission of a bioelectric signal called a nerve impulse. Neurons have two physiological properties: excitability, i.e. the ability to respond to stimuli and convert them into nerve impulses, and conductivity, i.e. the ability to transmit impulses. In formal neural networks, the behavior of biological neurons is imitated by a mathematical function, called activation, which has the property of being non-linear (to be able to transform the input in a useful way) and preferably of being differentiable (to allow learning by backpropagation of the gradient).Models closer to the biological neuron exist, in which the neuron emits impulses (or trains of impulses) with a certain frequency that depends on the incoming impulses. The activation function in this case can be represented as a function giving the variation of the average frequency of impulse emission with the input current.

[0053] More precisely, each layer 14 comprises neurons 16 taking their inputs from the outputs of the neurons 16 of the previous layer 14.

[0054] In the case of Figure 2, the neural network 12 described is a network comprising a single hidden layer of neurons 18. However, this number of hidden layers of neurons is not limiting.

[0055] The uniqueness of the hidden layer of neurons 18 means that the neural network 12 has an input layer 20 followed by the hidden layer of neurons 18, itself followed by an output layer 22.

[0056] The layers are indexable by an integer index i, the first layer corresponding to the input layer 20 and the last to the output layer 22.

[0057] Each layer 14 is connected by a plurality of synapses 24.

[0058] In biology, a synapse refers to a functional contact zone established between two neurons 16. Depending on its behavior, a biological synapse can excite or inhibit the downstream neuron in response to the upstream neuron. In formal neural networks, a positive synaptic weight corresponds to an excitatory synapse, while a negative synaptic weight corresponds to an inhibitory synapse. Biological neural networks learn by modifying synaptic transmissions throughout the network. Similarly, formal neural networks can be trained to perform tasks by modifying synaptic weights according to a learning rule. For the purposes of this application, a synapse 24 is a component performing a function equivalent to a synaptic weight of modifiable value.

[0059] A synaptic weight is therefore associated with each synapse 24. For example, it is a real number, which takes positive as well as negative values.

[0060] For each layer 14, the input of a neuron 16 is the weighted sum of the outputs of the neurons 16 of the previous layer 14, the weighting being done by the synaptic weights.

[0061] According to the example described, each layer 14 of neurons 16 is fully connected.

[0062] A fully connected neural layer is a layer in which the neurons of the layer are each connected to all the neurons of the previous layer. Such a type of layer is more often referred to as a

[0063] “fully connected”.

[0064] In this case, neural network 12 is a spiking neural network. A spiking neural network is often referred to by the acronym SNN, which refers to the English term for "Spiking Neural Network".

[0065] In such a spiking neural network, a neuron is a time-varying dynamic element as described above and characterized here by its pulse emission frequency. When a neuron 16, called pre-synaptic, upstream, emits an impulse, synapse 24 weights this impulse and transmits it to neuron 16, called post-synaptic, downstream, which eventually in turn emits an impulse.

[0066] The stimulation transmitted by synapse 24 is a stimulation of a part of neuron 16 downstream, called the membrane, which presents a potential. If this membrane potential charges beyond a threshold called activation, neuron 16 emits an impulse.

[0067] Specifically, synapse 24 performs a multiplication between the weight and the input activation. The input activation of downstream neuron 16 is the output signal sent by upstream neuron 16.

[0068] Downstream neuron 16 increases its membrane potential, compares it to a threshold, and emits an output impulse when the membrane potential exceeds this threshold.

[0069] In some cases, an upstream neuron 16 is continuously activated (like an input neuron) in order to add biases to the membrane potential of the downstream neuron 16 that enrich the expressiveness of the function learned by the neural network 12. Such a neuron 16 is a “bias neuron”.

[0070] The operation just described is valid in the other direction.

[0071] More specifically, in the example of Figure 2, for all layers 14 of neurons 16, except the first layer 20 and the bias neurons 16, the neurons 16 are connected by a synapse 24 which is bidirectional.

[0072] Therefore, the same calculation can be performed by exchanging the role of the two neurons 16.

[0073] As seen in Figure 1, the neuromorphic circuit 10 comprises a component 26, a configuration unit 28 of the component 26 and an interrogation unit 30 of the component 26.

[0074] In the example described, component 26 is unique but configurations could be envisaged where component 26 is not unique (see the paragraph “other particular examples”).

[0075] The component 26 can be excited according to a plurality of eigenexcitation modes with a respective population.

[0076] In the proposed schematic example, it is assumed that component 26 has 7 excitation modes.

[0077] Excitation modes or excited states are quantized. This means that excitation modes can be referenced by an integer i, with excitation modes usually being ordered by increasing energy.

[0078] In this case, as a non-limiting illustration, it is assumed that there are 7 excitation modes and that the integer i varies between 1 and 7.

[0079] The existence of excitation modes also makes it possible to define a reciprocal space in which the excitation modes are characterized by a respective eigenvector. Therefore, in the following, each excitation mode is denoted k, in reference to its eigenvector.

[0080] Furthermore, the population of an excitation mode k can be represented by the chemical potential ni.

[0081] Therefore, an excitation mode k with its respective population can be noted k(ni).

[0082] According to the example described, component 26 has two regimes, namely a linear regime (first regime) and a non-linear regime (second regime).

[0083] In the linear regime, the amplitude of the coupling between each pair of eigenexcitation modes ki to k7 is independent of the population of each of the two eigenexcitation modes ki to k7.

[0084] Thus, the population of excitation modes ki to k7 in the linear regime does not affect the energy of excitation modes ki to k7. As a result, excitation modes k to k7 can be considered quasi-orthogonal in reciprocal space.

[0085] When component 26 is in the linear regime that populates these excitation modes k to k7, this does not affect their energy and the excitation modes k to k7 can be considered as mainly orthogonal in reciprocal space. In this regime, the coupling between excitation modes k to k7 is thus weak.

[0086] In contrast, in the nonlinear regime, the amplitude of the coupling between each pair of eigenexcitation modes k to k7 depends on the population of each of the two eigenexcitation modes k to k7.

[0087] Of course, this is not limiting, the coupling amplitude can also be affected by the population in another excitation mode k to k7.

[0088] This corresponds to a regime in which the coupling between pairs of excitation modes k to k7 becomes significant, allowing the transfer of information from one excitation mode k to k7 to another excitation mode k to k7.

[0089] The coupling amplitudes can be formalized as a matrix A of synaptic dynamic weights. Each element of the matrix A is an ay element where the indices i and j refer to the two excitation modes k and kj, respectively. Each ay element depends on the population of excitation modes k and kj. The transition from the linear regime to a nonlinear regime is obtained by increasing the population of excitation modes ki to k7.

[0090] The component 26 described thus presents several excitation configurations, each excitation configuration being defined by the specific excitation modes ki to k7 excited and their respective population and corresponding to operation in the non-linear regime.

[0091] In other words, an excitation configuration corresponds to a set of ki(ni) values ​​for all values ​​of i.

[0092] This allows us to create a neural network with, on the one hand, the excitation modes ki to k7 and, on the other hand, the couplings.

[0093] Each excitation mode ki to k7 corresponds to a neuron and each coupling between two excitation modes ki to k7 corresponds to a synapse between the two neurons associated with the two excitation modes ki to k7 considered.

[0094] In other words, the neurons of the neural network 12 are realized by the specific excitation modes ki to k7 of the component 26 while the synapses are realized by the non-linear coupling matrix of the excitation modes ki to k7 of the component 26 (matrix A presented previously).

[0095] Component 26 is thus a programmable component capable of implementing the architecture of numerous neural networks.

[0096] The term "architecture" means the structure of the neural network 12, that is to say in particular, the position of each layer of neurons, the number of layers, the number of neurons in each layer and the links (synapses) between each neuron.

[0097] It is the configuration unit 28 of the component 26 which is suitable for programming the component 26.

[0098] The configuration unit 28 of the component 26 is capable of exciting the component 26 to obtain an excitation configuration chosen according to a desired neural network architecture.

[0099] The desired architecture is assumed to be known, knowing that it can in particular be obtained by carrying out training for a predetermined task or chosen by an expert as particularly appropriate for said task.

[0100] The desired architecture makes it possible to determine a matrix A which corresponds to values ​​of k(ni) for all values ​​of i, i.e. an excitation configuration.

[0101] This implies that the configuration unit 28 is capable of exciting the component 26 in the non-linear regime. According to a particular embodiment, the configuration unit 28 of the component 26 is capable of exciting the specific excitation modes ki to k7 according to an initial configuration different from the chosen excitation configuration.

[0102] Component 26 then naturally relaxes towards an equilibrium configuration, this equilibrium configuration being the chosen excitation configuration.

[0103] More formally, the configuration unit 28 carries out an initial configuration kio(n io) then component 26 can relax until it reaches a quasi-dynamic equilibrium corresponding to ki(ni) which is the desired configuration.

[0104] The initial configuration is chosen to be simpler to obtain, for example because it involves populating fewer different excitation modes ki to k7.

[0105] The query unit 30 is used to query the component 26, i.e. to send input data to the neural network to obtain at least one output data.

[0106] For this, according to the example of figure 1, the interrogation unit 30 comprises two sub-units 32 and 34.

[0107] The first subunit 32 is capable of selectively modifying the populations of first proper excitation modes ki to k7, the first proper excitation modes ki to k7 being the excitation modes of the input neurons (those of the input layer 20).

[0108] This means that the first subunit 32 is a subunit for sending input data to the component 26.

[0109] The second subunit 34 is suitable for measuring the populations of second proper excitation modes ki to k7, the second proper excitation modes ki to k7 being the excitation modes of the output neurons (those of the input layer 22).

[0110] This means that the second subunit 34 is a subunit for obtaining at least one output data from the component 26.

[0111] The operation of the neuromorphic circuit 10 is now described with reference to figure 3 which corresponds to a flowchart of an example of implementation of an inference method.

[0112] This method comprises three steps: a configuration step 40, a modification step E42 and a measurement step E44.

[0113] During the configuration step E40, the configuration unit 28 excites the component 26 to obtain a chosen excitation configuration.

[0114] For example, the configuration unit 28 receives a control signal indicating to it the excitation signal to be applied (for example, a radiofrequency signal with multiple frequencies as proposed in the embodiment of the following paragraph). This excitation signal has been previously calculated, for example by a simulation tool, to allow a desired neural network architecture to be physically produced.

[0115] As explained previously, this excitation signal can be used to obtain an initial configuration of the component 26 which will relax towards a final configuration corresponding to the desired architecture.

[0116] During the modification step E42, the interrogation unit 30 selectively modifies populations of first natural excitation modes.

[0117] This selective modification corresponds to the fact that it is desired that the neural network performs an inference on an input data.

[0118] The input data is converted into an excitation signal, for example by the aforementioned simulation tool, and a control unit sends a command indicating to the interrogation unit 30 that it must send said excitation signal.

[0119] Component 26 then evolves naturally until it reaches a stable state after such excitation.

[0120] During the measurement step E44, the interrogation unit 30 measures the populations of second natural excitation modes.

[0121] Thus, the interrogation unit 30 obtains the output data of the neural network, i.e. its prediction for the input data injected during the modification step E42.

[0122] This output data is obtained by converting the signal measured by the interrogation unit 30 by the inverse operation of that which was used to obtain the excitation signal for the input data.

[0123] In this way, an inference of the desired neural network is carried out in three stages.

[0124] Of course, it is possible to carry out these steps independently. In particular, it is possible to implement the physical realization of the neural network (configuration step 40) and the inference itself (steps 42 and 44).

[0125] In each case, the present neuromorphic circuit 10 corresponds to a physical realization of a neural network which is qualitatively different from all physical realizations known to date. Instead of designing and structuring individual nonlinear elements (neurons) and their interconnections (synapses) in real space, it is proposed to realize these elements in a high-dimensional reciprocal space. For this, it is sufficient to populate the eigenexcitation modes of the component 26 in the presence of strong nonlinearities to couple these excitation modes.

[0126] Thus, the connectivity of the neural network 12 can potentially be increased by several orders of magnitude, since it can utilize the full spectrum of eigenmodes in a given system. Values ​​as large as 10 6 can be obtained.

[0127] In other words, the physical realization in real space of components connected by physical links is no longer necessary here to obtain a large number of neurons presenting a large number of connections while guaranteeing good reconfigurability of synaptic weights.

[0128] The corollary of this observation is that neuromorphic circuit 10 is a more energy-efficient neuromorphic circuit.

[0129] It is worth noting here that a neuromorphic circuit 10 realizing highly interconnected neurons can be used for solving problems beyond neuromorphic computing. In particular, it could be envisaged to use this neuromorphic circuit 10 to solve difficult non-deterministic polynomial-time problems. Such problems are often referred to as "NP-hard". An example is solving a problem that can be projected onto an Ising model. Such potential could make such a neuromorphic circuit 10 a serious competitor to quantum computer realizations.

[0130] Furthermore, the neuromorphic circuit 10 is easily reconfigurable. The same neuromorphic circuit 10 can be used to perform different operations. This is achieved by simply modifying the signals of the configuration unit 28 and the interrogation unit 30 according to the intended application.

[0131] SPECIFIC EXAMPLE

[0132] A particular example of implementation of the neuromorphic circuit 10 is now described with reference to Figure 4.

[0133] In this specific example, component 26 is made of a ferromagnetic element.

[0134] According to the proposed example, the ferromagnetic element is yttrium iron garnet.

[0135] Yttrium iron garnet is more commonly referred to as YIG and refers to the element with the chemical formula YsFesO^. The abbreviation YIG here refers to the corresponding English name for "Yttrium Iron Garnet".

[0136] Component 26 is a thin layer here.

[0137] Such a layer is obtained by growing a layer or a multilayer of which at least one of the layers is magnetic.

[0138] The size and geometry of component 26 are chosen to have an appropriate energy separation between modes. Typically, for a single layer of YIG forming a cylinder, the size of the base of the cylinder is a few micrometers.

[0139] The shape of the base is circular (as seen in Figure 4) or rectangular.

[0140] The thickness of the layer forming component 26 is typically a few tens of nanometers.

[0141] Due to these dimensions, component 26 is a magnetic microstructure.

[0142] The component 26 is possibly covered with an overlayer 50 as is the case in the example illustrated.

[0143] The overlayer 50 serves to dynamically adjust the effective magnetic damping of the component 26. The damping is, by definition, the characteristic lifetime of the different excitation states.

[0144] Overcoat 50 is, for example, made of platinum.

[0145] The overlay 50 is controlled by a current generator. Such a generator can deliver direct current or a current forming time slots.

[0146] In such a system, collective oscillations of the magnetization around the equilibrium configuration are observed.

[0147] These oscillations are spin waves that exhibit quantized excitation modes.

[0148] The excitation modes therefore correspond here to a discrete frequency spectrum.

[0149] As explained in the previous section, these excitation modes can be used to realize the neurons of the neural network.

[0150] The various interactions occurring naturally in ferromagnetic elements can be described by an appropriate effective field, which itself depends on the magnetization, making the dynamics of the excitation modes, for sufficiently large deviations from the fundamental excitation mode, strongly non-linear.

[0151] More specifically, magnetization dynamics coupled with geometric effects and external stimuli such as applied fields and currents generate nonlinear exchange interactions and dipole-dipole interactions.

[0152] This corresponds to a regime in which most of the spin wave excitation modes are coupled together and the amplitude of the coupling depends on the population of each excitation mode.

[0153] As explained in the previous section, these couplings can be used to realize the synapses of the neural network benefiting from high connectivity. In addition, it is possible to dynamically modify the synaptic weights of each synapse by controlling the population of each of the spin wave modes by applying an external signal.

[0154] For this purpose, an antenna is used as a configuration unit 28.

[0155] The configuration unit 28 is suitable for revealing non-linearities at very low excitation amplitude, typically for a few microTeslas (-30 dBm) of alternating magnetic fields at 10 GHz.

[0156] Stronger excitation of component 26 allows reaching deep nonlinear regimes of spin wave dynamics. In such a regime, indexing linear modes is no longer an adequate physical description of the system's behavior.

[0157] The configuration unit 28 then uses radio frequency signals in the time and frequency domains to selectively populate the specific excitation modes, the spectral density of these signals respecting the previous excitation conditions.

[0158] The interrogation unit 30 is either the same antenna as the configuration unit 28 or, as shown in Figure 4, another antenna.

[0159] It is worth noting that YIG is particularly suitable here because thin films can be nanostructured down to 300 nanometers (nm) without any negative impact on their magnetic properties. In particular, YIG exhibits very low spin wave damping, with this damping being as high as 8 x 10 -5 in the geometries described in this section.

[0160] This low damping gives rise to an extremely rich dynamic behavior with numerous non-linear effects, thus making the excitation modes accessible. Good exploitation of these non-linear effects makes it possible to easily control the kj(ni) configurations.

[0161] However, other materials than YIG can be considered.

[0162] In particular, the materials are metallic alloys of the elements, such as NiFe, CoFeB or CoNi.

[0163] Alternatively, the materials are Heusler compounds such as CoFeMnSi.

[0164] It is also possible to use doped YIG or to substitute the yttrium Y in YIG with another atom, for example thulium (Tm) or bismuth (Bi).

[0165] The neuromorphic circuit 10 is here a spintronic device. As such, it has the advantage of being compatible with CMOS circuits. OTHER SPECIFIC EXAMPLES

[0166] The particular realization described in the previous section is not limiting since many other physical systems exhibit analogous behaviors of their excitation spectrum.

[0167] In particular, photonic crystals, which are optical cavities made by nanostructuring, are multimode elements. In addition, the nonlinear regime and the couplings between modes can be obtained by modulation of the dielectric constant by generation of electron-hole pairs by optical pumping.

[0168] As another example, we can consider metamaterials, which are versatile systems where resonances can be defined geometrically. By using low-dissipative materials, nonlinear operation would be achieved. A superconducting resonator is an example of such a metamaterial.

[0169] More generally, it is sufficient to have a physical system capable of being excited according to a plurality of specific excitation modes with a respective population and that each pair of excitation modes is coupled by a respective and controllable coupling.

[0170] It should be noted here that it is not necessary to be able to control all the eigenexcitation modes of the physical system but only a part and that on this part, the coupling is controllable by the population of these eigenexcitation modes.

[0171] Furthermore, inaccessible modes can nevertheless be used to increase the number of degrees of freedom and thus significantly increase the depth of the resulting neural network.

[0172] Such control may use any physical effect so that the configuration 28 and interrogation 30 units may be an optical excitation unit, an electrical excitation unit.

Claims

CLAIMS 1. Neuromorphic circuit (10) physically realizing a neural network, the neural network comprising a set of neurons connected by synapses, the neuromorphic circuit (10) comprising: - at least one component (26) capable of being excited according to a plurality of specific excitation modes with a respective population, the component (26) thus being capable of presenting several excitation configurations, each excitation configuration being defined by the specific excitation modes excited and their respective population, each pair of excitation modes being coupled by a respective coupling, each specific excitation mode being a neuron of the neural network and each coupling being a synapse of the neural network, - a configuration unit (28) of the at least one component (26), the configuration unit (28) being capable of exciting the component (26) to obtain an excitation configuration chosen according to a desired neural network architecture, the desired architecture comprising input neurons and output neurons, and - an interrogation unit (30) of the at least one component (26), the interrogation unit (30) of the component (26) being capable of selectively modifying the populations of first specific excitation modes and of measuring the populations of second specific excitation modes, the first specific excitation modes being the excitation modes of the input neurons and the second specific excitation modes being the excitation modes of the output neurons.

2. Neuromorphic circuit according to claim 1, in which the component (26) is capable of being excited in a regime in which the amplitude of a coupling between each pair of eigenexcitation modes in each excitation configuration depends on the respective populations of each of the two eigenexcitation modes.

3. Neuromorphic circuit according to claim 1 or 2, wherein the at least one component (26) has two regimes, a first regime in which the amplitude of the coupling between each pair of eigenexcitation modes is independent of the population of each of the two eigenexcitation modes and a second regime in which the amplitude of the coupling between each pair of eigenexcitation modes depends of the population of each of the two specific excitation modes, the configuration unit (28) being capable of exciting the component (26) in the second regime.

4. Neuromorphic circuit according to any one of claims 1 to 3, in which the configuration unit (28) is capable of exciting the specific excitation modes according to an initial configuration different from the chosen excitation configuration, the component (26) relaxing towards the chosen excitation configuration.

5. Neuromorphic circuit according to any one of claims 1 to 4, in which the at least one component (26) is a layer of ferromagnetic element.

6. The neuromorphic circuit of claim 5, wherein the ferromagnetic element is an yttrium iron garnet.

7. Neuromorphic circuit according to any one of claims 1 to 4, in which the at least one component (26) is chosen from the list consisting of: - a magnetic microstructure - a cavity, - a metamaterial, and - a superconducting resonator.

8. Neuromorphic circuit according to any one of claims 1 to 7, in which the configuration unit (28) is chosen from: - an optical excitation unit, and - an electrical excitation unit.

9. Neuromorphic circuit according to any one of claims 1 to 8, in which the interrogation unit (30) is chosen from: - an optical excitation unit, and - an electrical excitation unit.

10. Method for physically producing a neural network having a desired architecture, the neural network comprising a set of neurons connected by synapses, the physical production method comprising: - a step of configuring at least one component (26) of a neuromorphic circuit (10) physically realizing a neural network having been configured, the at least one component (26) being capable of being excited according to a plurality of specific excitation modes with a respective population, the component (26) thus being capable of having several excitation configurations, each excitation configuration being defined by the excited specific excitation modes and their respective population, each pair of excitation modes being coupled by a respective coupling, each specific excitation mode being a neuron of the neural network and each coupling being a synapse of the neural network, the configuration step being implemented by a configuration unit (28) of the at least one component (26), the configuration unit (28) being part of the neuromorphic circuit (10) and comprising the excitation of the at least one component (26) to obtain an excitation configuration chosen according to the desired architecture.

11. A method for inferring a neural network having a desired architecture, the neural network comprising a set of neurons connected by synapses, at least one component (26) of a neuromorphic circuit (10) physically implementing a neural network having been configured, the at least one component (26) being capable of being excited according to a plurality of specific excitation modes with a respective population, the component (26) thus being capable of having several excitation configurations, each excitation configuration being defined by the excited specific excitation modes and their respective population, each pair of excitation modes being coupled by a respective coupling, each specific excitation mode being a neuron of the neural network and each coupling being a synapse of the neural network, the configuration having been implemented by a configuration unit (28) of the at least one component (26),the configuration unit (28) forming part of the neuromorphic circuit (10) and comprising the excitation of the at least one component (26) to obtain an excitation configuration chosen according to the desired architecture, the inference method comprising:, - a step of selectively modifying the populations of first proper excitation modes, the first proper excitation modes being the excitation modes of the input neurons, the modification step being implemented by an interrogation unit (30) of the neuromorphic circuit (10), and - a step of measuring the populations of second proper excitation modes, the second proper excitation modes being the excitation modes of the output neurons, the measurement step being carried out after reaching a stable state for the at least one component (26), the measuring step being implemented by the interrogation unit (30).

Images