Neuromorphic component using countable magnetic textures

Abstract

The invention relates to a neuromorphic component (28) physically implementing a neuron and comprising: - for each input value: - a track (32) having countable and displaceable local magnetic modifications, - a local magnetic modification control unit (40) causing the local magnetic modifications in a detection zone (46) common to all the tracks (32), and - a local magnetic modification detection unit applying an activation function to an electrical quantity corresponding to the number of local magnetic modifications detected in the detection zone (46) so as to obtain the output value, the detection unit (36) being placed on the detection zone (46).

Description

[0001] Neuromorphic component using countable magnetic textures

[0002] The present invention relates to a neuromorphic component physically realizing a neuron of a neural network and a neuromorphic circuit comprising such a component.

[0003] 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.

[0004] 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.

[0005] 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 many 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.

[0006] 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.

[0007] 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.

[0008] 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.

[0009] 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.

[0010] A neural network based on optical technologies is also known.

[0011] 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.

[0012] 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.

[0013] There is therefore a need for a neuromorphic component capable of implementing neurons in a neural network having greatly improved compactness with good computing performance. To this end, the description describes a neuromorphic component physically realizing a neuron of a neural network connected to at least one synapse, each synapse carrying a respective weight, the neuron taking as input at least one input value weighted by a respective weight to obtain an output value, the neuromorphic component comprising:

[0014] - for each input value:

[0015] - a track capable of presenting countable and movable local magnetic modifications,

[0016] - a unit for controlling local magnetic modifications, the control unit being capable of annihilating and creating local magnetic modifications, the control unit also being capable of bringing the local magnetic modifications into a zone for detecting local magnetic modifications, the detection zone being common to all the tracks, the control unit comprising a sub-unit for applying electrical pulses, the application sub-unit being capable of applying a current proportional to the input value, and a sub-unit for modifying at least one magnetic property of the track with an amplitude proportional to the weight weighting the input value, and

[0017] - a unit for detecting local magnetic modifications, the detection unit being capable of applying an activation function to an electrical quantity corresponding to the number of local magnetic modifications detected in the detection zone to obtain the output value, the detection unit being positioned on the detection zone.

[0018] According to particular embodiments, the neuromorphic component has one or more of the following characteristics, taken in isolation or in all technically possible combinations:

[0019] - the detection unit uses a physical effect to achieve the activation function, the effect being a magnetotransport effect.

[0020] - the detection unit is a magnetic junction.

[0021] - the tracks are separated.

[0022] - the detection zone is a reservoir common to all tracks.

[0023] - the neuromorphic component further comprises a reinjection unit, the reinjection unit being capable of reinjecting the output value into other neurons of the neural network.

[0024] - local magnetic modifications are chosen from the list consisting of: - skyrmions,

[0025] - hopfions,

[0026] - bubbles,

[0027] - blood cells,

[0028] - cocoons,

[0029] - strands, and

[0030] - domain walls.

[0031] - the detection zone is made of a material being:

[0032] - an element selected from Ni, Fe, Co, Ge, B, C, P, S, Mn, Cr, I, Br, Si, Pt, Al, Ga, Ta, Ru, Ht, Mg, Bi, Mo, Te, Se, W, N, Ti, Ir, V, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Er, Tm, and Yb,

[0033] - an alloy comprising one or more of the aforementioned elements, and

[0034] - an oxide comprising one or more of the above-mentioned elements.

[0035] - each track is made from a material chosen from:

[0036] - an element chosen from 10. Ni, Fe, Co, Ge, B, C, P, S, Mn, Cr, I, Br, Si, Pt, Al, Ga, Ta, Ru, Ht, Mg, Bi, Mo, Te, Se, W, N, Ti, Ir, V, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Er, Tm and Yb

[0037] - an alloy comprising one or more of the aforementioned elements, and

[0038] - an oxide comprising one or more of the above-mentioned elements.

[0039] - the modification subunit uses a physical effect chosen from an electrostatic effect, a magneto-ionic effect, an electro-ionic effect, an electrochemical effect, a piezoelectric effect and a magnetostrictive effect.

[0040] - the application subunit is capable of applying a current to the track.

[0041] - the application sub-unit is capable of applying a current in a layer above or below the track. Such an application is suitable for the case where the track is made of an insulating material.

[0042] - the application sub-unit is capable of generating a current in another track arranged so as to heat the track by the Joule effect. This Joule heating causes the formation of local magnetic modifications (for example nucleation of skyrmions) on said track.

[0043] - the number of magnetic modifications of each track is proportional to the product of the synaptic weight and one or more of the characteristics of the electrical pulses (current pulse.

[0044] - the characteristic of the current pulse is chosen from the group consisting of the current amplitude, the pulse duration and the number of pulses. The description also relates to a neuromorphic circuit physically realizing a neural network, the neuromorphic circuit comprising a set of neuromorphic components as previously described.

[0045] According to a particular embodiment, the set of neuromorphic components is arranged to form a deep network.

[0046] In this description, the expression "suitable for" means indifferently "adapted for", "adapted to" or "configured for".

[0047] 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:

[0048] - Figure 1 is a schematic representation of an example of a part of a neuromorphic circuit comprising a plurality of neurons and synapses,

[0049] - Figure 2 is a schematic representation of an example of a neural network,

[0050] - Figure 3 is a schematic representation of an example of a neuromorphic component physically realizing a neuron of a neural network, the neuromorphic component being seen from above,

[0051] - Figure 4 is a schematic side view representation of the neuromorphic component of Figure 3,

[0052] - Figure 5 is a schematic representation of part of the operation of the neuromorphic component of Figure 3,

[0053] - Figure 6 is a schematic representation of another example of a neuromorphic component physically realizing a neuron of a neural network, the neuromorphic component being seen from above, and

[0054] - Figure 7 is a schematic representation of an example of a neuromorphic circuit comprising neuromorphic components according to Figure 3.

[0055] A portion of a neuromorphic neural network circuit 10 is shown in Figure 1.

[0056] The neuromorphic circuit 10 is suitable for implementing a neural network 12 as shown diagrammatically in Figure 2.

[0057] 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 preceding layer 14. By definition, in biology, a neuron, or a 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, that is, the ability to respond to stimulation and convert it into nerve impulses, and conductivity, that is, the ability to transmit impulses.

[0058] 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 derivable (to allow learning by backpropagation of the gradient).

[0059] An activation function (sometimes called a "simulation of electrical transfer between synapses") is a mathematical function that occurs in all types of neurons in a neural network. The "square root" function, linear rectification (more often referred to by the acronym ReLu, which refers to the corresponding English term for "Rectified Linear Unit"), sigmoid, or hyperbolic tangent are examples of functions that can be used as an activation function.

[0060] For the purposes of this application, a neuron 16 is a component performing a function equivalent to these latter models.

[0061] More precisely, each layer 14 comprises neurons 16 taking their inputs from the outputs of the neurons 16 of the previous layer 14.

[0062] 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.

[0063] The uniqueness of the hidden layer of neurons 18 means that the neural network 10 has an input layer 20 followed by the hidden layer of neurons 18, itself followed by an output layer 22.

[0064] Each layer 14 is connected by a plurality of synapses 24.

[0065] 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.

[0066] 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.

[0067] 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.

[0068] According to the example described, each layer 14 of neurons 16 is fully connected.

[0069] A fully connected layer of neurons is one in which the neurons in that layer are each connected to all the neurons in the previous layer. This type of layer is more often referred to as "fully connected."

[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] The neuromorphic circuit 10 comprises a set of neuromorphic components 26.

[0074] By definition, a neuromorphic component is a component capable of performing a function in a neural network, in particular neuromorphic components performing synapse functions and neuron functions respectively.

[0075] According to the example, the neuromorphic circuit 10 comprises neuromorphic components performing the function of a neuron 28 and neuromorphic components performing the function of a synapse 30.

[0076] The neuromorphic components 26 are arranged to form a neural network, preferably a deep neural network.

[0077] In the remainder of the description, to clarify the point, the elements performing one of the aforementioned functions are designated by the name of the function and not by a formulation of the type “component / circuit performing function X” or “comprising / circuit physically implementing function X”.

[0078] Instead of the expression "neuromorphic component performing the function of a neuron 28", the expression "neuron 28" will be used and, instead of the expression "neuromorphic component performing the function of a synapse 30", the expression "synapse 30" will be used.

[0079] An example of neuron 28 is shown in Figure 3

[0080] As explained previously, neuron 28 takes as input at least one input value weighted by a respective weight to obtain an output value.

[0081] In the following, it is assumed that neuron 28 takes n values ​​as input, where n is an integer greater than 1.

[0082] The number n is, for example, greater than 10, preferably greater than 50 and more preferably greater than 100.

[0083] This means that neuron 28 is connected to n different synapses 30.

[0084] To implement such a function physically, the neuron 28 comprises n tracks 32, n control units 34 and a detection unit 36.

[0085] The neuron 28 thus comprises, for each input value, a track 32 and a respective control unit 34.

[0086] Each track 32 is capable of presenting countable and movable local magnetic modifications.

[0087] A local magnetic modification 40 is a non-collinear magnetic texture having particle or soliton properties, and which can be counted.

[0088] These textures are also movable textures, that is, they can be moved in a controlled manner within neuron 28.

[0089] Furthermore, in this case, the textures have a characteristic dimension of less than one micron, so that the textures are submicron in size.

[0090] As an illustration, a typical example of local magnetic modifications 40 is a set of skyrmions, which correspond to a vortex or spin vortex on a surface.

[0091] Alternatively, the local magnetic modifications 40 are hopfions, bubbles, globules, cocoons, strands or domain walls.

[0092] In the following, the term “magnetic particles 40” will be used to designate these local magnetic modifications.

[0093] The tracks 32 can here be seen as injection tracks or injectors of magnetic particles 40 depending on the operating mode of the neuron 28 as will be explained later (creation and movement of the particles 40 compared to movement only of the particles 40).

[0094] According to the example of Figure 3, the tracks 32 are rectilinear elements.

[0095] Each track 32 extends mainly along a main direction noted as the main direction X. The track 32 typically has a length of around ten micrometers (dimension along the main direction X) and a width of a few micrometers (dimension along a transverse direction noted as Y).

[0096] To complete the identification, a stacking direction of the layers is also defined, this direction being noted as stacking direction Z.

[0097] Additionally, according to the illustrated geometry, the tracks 32 are parallel along the main direction X.

[0098] However, this geometry is not limiting, other shapes being conceivable for the tracks 32, in particular curved shapes and other relative arrangements of the tracks 32 (non-parallel) as long as the tracks 32 remain separate.

[0099] Each track 32 is made of a material which comprises an element chosen from Ni, Fe, Co, Ge, B, C, P, S, Mn, Cr, I, Br, Si, Pt, Al, Ga, Ta, Ru, Hf, Mg, Bi, Mo, Te, Se, W, N, Ti, Ir, V, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Er, Tm and Yb.

[0100] Depending on the case, the material is the element or an alloy comprising one or more of these elements or an oxide comprising one or more of these elements.

[0101] Furthermore, each track 32 is separated from another track 32 by an intertrack space 48 which can be made with similar materials.

[0102] The 32 tracks are, for example, obtained by etching in a magnetic multilayer.

[0103] As a particular illustration, a lithography operation could be carried out on the n parallel tracks 32 in a metallic magnetic multilayer film allowing the generation of skyrmions. Ta5 / Co2oFe6oB2o / Tao.o8 / MgO is a particular example of such a film.

[0104] The particles 40 circulating in each track 32 as well as the magnetic properties of each track 32 are controlled by a control unit 34 which is specific to each track 32.

[0105] Each control unit 34 comprises an application sub-unit 42 and a modification sub-unit 44.

[0106] The control unit 34 is a magnetic particle control unit 40.

[0107] In this respect, the control unit 34 is capable of annihilating and creating the magnetic particles 40.

[0108] The control unit 34 is also capable of bringing the magnetic particles 40 into a detection zone 46 of these magnetic particles 40.

[0109] These annihilation, creation and displacement operations are notably achievable by the application sub-unit 42 as will be explained later in this description. The application sub-unit 42 is an application sub-unit 42 of electrical pulses.

[0110] The application sub-unit 42 is thus capable of applying a current to the track 32 to perform operations on the magnetic particles 40.

[0111] In particular, the application sub-unit 42 is capable of applying a current proportional to the input value associated with the track 32 considered.

[0112] In this example, the detection area 46 is the oval area surrounded by dotted lines.

[0113] This detection zone 46 comprises a part of all the tracks 32 and the part of the intertrack 48 located between these parts.

[0114] The detection zone 46 is thus common to all the tracks 32, that is to say that the detection zone 46 is shared by each track 32.

[0115] Due to the materials mentioned above, the detection zone 46 is here made of several materials, each material comprising an element chosen from Ni, Fe, Co, Ge, B, C, P, S, Mn, Cr, I, Br, Si, Pt, Al, Ga, Ta, Ru, Hf, Mg, Bi, Mo, Te, Se, W, N, Ti, Ir, V, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Er, Tm and Yb. Depending on the case, each material is the element or an alloy comprising one or more of these elements or an oxide comprising one or more of these elements.

[0116] The control unit 34 is also capable of controlling the properties of the track 32 by using the modification sub-unit 44.

[0117] Indeed, the modification sub-unit 44 is capable of modifying one (or more) of the magnetic properties of the track 32 with an amplitude proportional to the weight weighting the input value.

[0118] A first example of a magnetic property that can be modified by modification subunit 44 is perpendicular magnetic anisotropy.

[0119] The energy of perpendicular magnetic anisotropy is expressed as K(S t ■ z) 2 with K the anisotropy and S L • z the out-of-plane component of the local magnetization.

[0120] A second example of a magnetic property that can be modified by the modification subunit 44 is the asymmetric exchange interaction.

[0121] Such an interaction can notably be described by the Dzyaloshinskii-Moriya interaction, itself described by the Hamiltonian H 0M[ Lj = ■ St x S7), the bold letters being vectors, S and Sj being two close spins and Dÿ is the corresponding Dzyaloshinskii-Moriya vector.

[0122] A third example of a magnetic property that can be modified by the modification subunit 44 is the symmetric exchange interaction.

[0123] Such an interaction can be written by the Heisenberg Hamiltonian. The perpendicular magnetic anisotropy and the symmetric and antisymmetric exchange interactions are governed by the nature of the atoms, their positions relative to each other and those of their electron cloud. These properties condition the possibility of skyrmion nucleation and the quantity of skyrmions that will be nucleated.

[0124] The modification here is a non-volatile and reversible modification using a gate voltage.

[0125] This makes it possible to obtain a linear variation in the quantity of magnetic particles 40 created or displaced in the detection zone 46.

[0126] For this, according to a particular example, electrodes are deposited locally on each track 32.

[0127] Such electrodes can be made of oxide with a material such as HfOs or MgO.

[0128] The electrodes are further placed near the detection zone 46.

[0129] To achieve such a modification, modification subunit 44 uses a physical effect.

[0130] For example, the effect is chosen from the following effects: electrostatic effect, electroionic effect, magnetoionic effect, electrochemical effect, piezoelectric effect and magnetostrictive effect.

[0131] In particular, it may be envisaged to use the reversible migration of ions using an ionic liquid or the reversible migration of oxygen or hydrogen atoms from an oxidized layer deposited above the tracks 32 to modify the local magnetic properties. The chemical modifications generated by this ionic displacement (in oxide layers or in the ionic liquid) lead to a modification of the orbitals and / or the state densities of the ferromagnetic layers resulting in a residual change in the magnetic properties.

[0132] During learning, the weights can thus be adjusted using an electrical stimulus, this stimulus coming from an additional generator.

[0133] Thus, at the level of neuron 28, the control units 34 allow electrical current pulses to be sent by the corresponding application sub-unit 42 in a track 32. The track 32 identified by the index i is thus crossed by a current pulse noted Ji.

[0134] In parallel, the modification sub-unit 44 modifies a magnetic property of the track 32 so that the number of magnetic particles 40 created and / or displaced is proportional to the product of the amplitude of the current pulse with the synaptic weight Wi. This corresponds well to the modulation of the inputs Ji by synaptic weights Wi which is sought here. The detection unit 36 ​​is positioned on the detection zone 46 as visible in FIG. 4.

[0135] The detection unit 36 ​​is capable of detecting magnetic particles 40.

[0136] The detection unit 36 ​​measures the number of magnetic particles 40 by means of an electrical reading.

[0137] The detection unit 36 ​​applies an activation function to the electrical quantity corresponding to the number of magnetic particles 40 to obtain a new electrical quantity, this new electrical quantity is the output value.

[0138] The activation function is non-strictly linear, i.e., non-linear and monotonic. Indeed, a linear activation function would not be suitable, because the synaptic weights are adjusted using the derivative of the activation function. If the activation function were linear, all weights would be adjusted identically, regardless of the inputs.

[0139] The detection unit 36 ​​thus has two roles: on the one hand, to count the magnetic particles 40 in the detection zone 46 and, on the other hand, to convert the number of magnetic particles 40 into an electrical output equal to the result of the application of the activation function to this number of magnetic particles 40.

[0140] For this purpose, the detection unit 36 ​​performs a non-linear electrical reading. Such a reading can be carried out in several ways.

[0141] In particular, the detection unit 36 ​​may be a spin valve. A spin valve is a component, comprising two or more layers of conductive magnetic materials, the electrical resistance of which can be varied between several values ​​depending on the relative angle between the magnetizations of the layers.

[0142] According to one example, the detection unit 36 ​​is thus a magnetic tunnel junction.

[0143] The magnetic tunnel junction corresponds to the oval area of ​​detection zone 46.

[0144] It is, for example, obtained by depositing CoFeB above the detection zone 46 described previously, which creates a tunnel junction with the MgO layer.

[0145] Such a tunnel junction has a resistance varying with the number of magnetic particles 40 positioned in the detection zone 46.

[0146] Thus, the tunnel junction uses a tunnel magnetoresistance effect. This magnetoresistive effect having a very high amplitude, this allows easy detection of magnetic particles 40.

[0147] Furthermore, the variation of the resistance of the tunnel junction is directly non-linear with the number of magnetic particles 40 in the detection zone 46 if the starting state is the most resistive state (case where the magnetization of the associated layers is antiparallel) and the magnetoresistance effect is sufficiently large. This results in better compactness.

[0148] In addition, the resistance of the tunnel junction is much greater than that of the 32 tracks. This makes it possible to limit the propagation of current between the different 32 tracks.

[0149] Finally, the materials generally used for tunnel junctions are very similar to those used for the magnetic multilayers forming the tracks 32 and the detection zone 46, so that manufacturing is easier (better compatibility between the materials).

[0150] As a particular example of materials, CoFeB and MgO can be cited to form a tunnel junction formed by alternating CoFeB / MgO / CoFeB.

[0151] According to another example, the detection unit 36 ​​uses a physical effect involving a linear variation of the resistance with the number of magnetic particles 40 detected.

[0152] The physical effect is, for example, an anisotropic magnetoresistance effect, an extraordinary Hall effect, a spin Hall effect, or a non-collinear magnetoresistance effect.

[0153] The detection unit 36 ​​is then a magnetic junction measuring the variation in the resistance of the track 32 as a function of the number of magnetic particles 40 present in the detection zone 46.

[0154] Alternatively, the effect may be an extraordinary Hall effect because the Hall voltage varies with the number of magnetic particles 40 present in the detection area 46.

[0155] The detection unit 36 ​​can also use a Nernst effect which corresponds to the fact that the measured transverse voltage is proportional to the number of magnetic particles 40 generated by a thermal current.

[0156] In such cases, the detection unit 36 ​​comprises an additional element for applying a subsequent non-linear function.

[0157] For example, such an additional element is a transistor. This transistor can be produced using CMOS technology, which is compatible with the materials used to produce the 32 tracks.

[0158] Alternatively, the additional element may take advantage of the change in size of the magnetic particles 40 depending on the magnetic properties of the track 32, this change in size also being able to be detected electrically.

[0159] It is also possible that the additional element uses a magneto-thermoelectric effect, this involving the measurement and / or application of temperature gradients.

[0160] To avoid the propagation of current between the different tracks 32, tracks 32 may be used whose transverse resistance (in the direction perpendicular to the main direction in which the track 32 extends) is greater than the resistance of the track 32 (defined as the resistance in the main direction).

[0161] It is also possible to consider using a graphene layer to achieve electrical detection.

[0162] Indeed, graphene exhibits electrical properties, including electrical transport properties, varying with the magnetic fields created by magnetic particles 40.

[0163] For this, at the detection zone 46, an insulating layer topped with the graphene layer is deposited. The insulating layer is thin enough so that the leakage fields created by the magnetic particles 40 are sufficient to be detected in the graphene, while making it possible to limit the disturbances of the magnetic particles 40 linked to the presence of the graphene.

[0164] The graphene layer thus has no electrical contact with the tracks 32, which prevents current from flowing from one track 32 to another.

[0165] In each case, the effects are of a magneto-transport type effect and this allows the detection unit 36 ​​to perform a non-linear operation corresponding to the activation function, a function which is necessary when a neuromorphic calculation is implemented.

[0166] The operation of the neuromorphic component 26 realizing the neuron 28 is now described through several operations.

[0167] As for the calculation operation, for a neuron 28 connected by n synapses 30, n currents of current density Ji generate in each track 32 a number of magnetic particles 40 Ni proportional to both the input value Ji and the synaptic weights Wi.

[0168] As explained previously, each synaptic weight Wi corresponds to a local modification of the magnetic properties which can be modulated in a reversible and non-volatile way by the gate voltage.

[0169] In other words, the current density inputs Ji are modulated by the synaptic weights Wi so that the number Ni of magnetic particles 40 in the detection area 46 of the track 32 i is proportional to Ji* Wi.

[0170] The detection unit 36 ​​directly performs the sum of the different numbers Ni of magnetic particles 40 to obtain the total number of magnetic particles 40, namely N = and applies a non-linear function f on this total number N to obtain an electrical output V out , so that V 0Ut = f(N) = f dNt) To obtain such a situation where Ni magnetic particles 40 circulate in a track 32 i, two cases are possible: nucleation of magnetic particles 40 with a displacement or displacement alone (from a reservoir).

[0171] Two concrete examples of the implementation of these two cases are visible in Figure 5.

[0172] Each example comprises two parts, a part on the left corresponding to the state of track 32 considered before the application of the current pulse Ji and a part on the right corresponding to the state of track 32 considered after the application of the current pulse Ji.

[0173] The example noted I corresponds to the case of nucleation of magnetic particles 40.

[0174] Before the application of a current pulse, as illustrated in the left part, no magnetic particles 40 are present in the track 32 i whereas after the application of a current pulse, as illustrated by the right part, Ni magnetic particles 40 are nucleated (and possibly then displaced) in the detection zone 46, this nucleation being proportional to both the amplitude of the current pulse Ji and the synaptic weight Wi.

[0175] Depending on the embodiments, the current density of the current pulse, the time of the current pulse, or the number of current pulses over a predefined time can be varied to encode the desired input value. This provides three degrees of freedom.

[0176] Alternatively or additionally, in the case where the magnetic particles 40 are skyrmions, a local increase in temperature allows skyrmions to nucleate. Also, in such a case, it may be considered to encode the input value by the power or energy density of a thermal current. These are two other degrees of freedom.

[0177] In this example I, the magnetic particles 40 are then removed from the tracks 32 at the end of each iteration of the computational operation.

[0178] The example noted II corresponds to the case of a displacement of magnetic particles 40 already created.

[0179] Before the application of a current pulse, as illustrated in the left part, the magnetic particles 40 are already present in the track 32 i, for example, in a dedicated area forming a reservoir.

[0180] After applying a current pulse, the right part shows that a portion of the magnetic particles 40 has been moved to the detection zone 46 and that another portion of the magnetic particles 40 has remained in the reservoir. The number of magnetic particles 40 moved is equal to Ni. In this example II, the magnetic particles 40 are then brought back to the reservoirs of the tracks 32 at the end of each iteration.

[0181] This analysis of the different cases shows that neuron 28 is also capable of performing a reset operation of the calculation operation.

[0182] For a reset operation consisting of removing magnetic particles 40, it may be considered to saturate the magnetization of the neuron 28 using an external magnetic field before use.

[0183] Such a magnetic field is then a field applied in a direction opposite to the magnetization of the center of the skyrmions, that is to say oriented according to the stacking direction Z.

[0184] Alternatively, it is possible to apply electrically controlled local magnetic fields in a predefined location and move the skyrmions to this predefined location.

[0185] It is also possible to use the magnetization of the upper layer of the magnetic tunnel junction.

[0186] For a reset operation consisting of returning the magnetic particles 40 to the reservoir, one (or more) current pulses may be injected into the track 32 having an amplitude and duration similar to the initial current pulse but with an inverse polarity.

[0187] It may be noted here that the synaptic weights may be used to increase or decrease the speed of the set of magnetic particles 40 and that only Ni magnetic particles 40 are in the detection area 46, the other particles 40 being around the detection area 46.

[0188] Alternatively, it is also possible to have a wider reservoir and a constriction, made by, for example, lithography at the same time as the reservoir itself, and that the current pulse serves only to move Ni magnetic particles 40 out of the reservoir.

[0189] In this case, neuron 28 is also capable of performing an initial tank filling operation.

[0190] During such an initialization operation, the magnetic particles 40 are, for example, created by the application of a current pulse having a large amplitude.

[0191] For non-volatile magnetic particles 40, this initial current pulse is sufficient to carry out all iterations of the calculation operation. It will be understood that the current pulses used during these iterations have a lower amplitude to avoid the generation of additional magnetic particles 40.

[0192] Alternatively, the magnetic particles 40 may be generated magnetically or electrically and then moved to the desired locations prior to use of the neuron 28.

[0193] Neuron 28 is also capable of implementing a learning operation.

[0194] The learning operation involves the use of the gate voltage, this voltage being adjusted according to the value of the synaptic weight to be coded.

[0195] Such voltage adjustment is carried out taking care not to disturb the skyrmions present in track 32 associated with the modification sub-module.

[0196] It is thus proposed a neuron 28 carrying out the weighted sum of the inputs and the synaptic weights, directly with the number of magnetic particles 40 injected, and to carry out the electrical reading of this sum thanks to a magnetoresistive effect, which intrinsically carries out a non-linear function.

[0197] Thus, the neuromorphic component 28 described makes it possible to carry out the storage of synaptic weights, the calculation of the weighted sum and the application of the activation function in an integrated manner in a single compact device.

[0198] Such a neuromorphic component 28 relies in particular on the possibility of electrically controlling the nucleation, displacement and annihilation of magnetic particles 40 in magnetic multilayers, in order to inject (nucleate, displace or both at the same time) a certain number of magnetic particles 40 proportional to the input and the synaptic weight for each track 32. The neuromorphic component 28 therefore benefits from the fact that the magnetic particles 40 can easily be controlled and counted.

[0199] The neuromorphic component 28 is, among other things, a spintronic device, so that this component has the same advantages as all spintronic components. In particular, the neuromorphic component 28 benefits from the nonvolatile character intrinsically linked to the magnetic properties of the magnetic particles 40.

[0200] Such a component is, moreover, compatible with CMOS components and other spintronic components.

[0201] The neuromorphic component 28 also exhibits reliable and reproducible operation over a wide frequency range. This reliability is accompanied by good resistance to electromagnetic radiation from the environment. Such an effect is amplified by the fact that the weighted sum involves statistics on a large number of magnetic particles 40. This also makes it possible to reduce the influence of the distribution of the sizes of magnetic particles 40, this distribution being able to generate noise.

[0202] Moreover, the operations performed by neuron 28 can be relatively rapid because the nucleation and displacement of magnetic particles 40 are physical effects exhibiting relatively rapid dynamics.

[0203] Moreover, this dynamic implies a low current density, which allows to reduce the energy consumed by the neuromorphic component 28.

[0204] The magnetic particles 40 are also submicron in size, so the dimensions of the tracks 32 are relatively small, which also limits electrical energy consumption.

[0205] This gain in consumption is reinforced in the case of displacement alone because the energy necessary for the displacement of the magnetic particles 40 is less than that necessary for their nucleation.

[0206] Magnetic particles 40 also have the advantage of being able to be nucleated, stabilized and controlled at room temperature and in the absence of an external magnetic field, as well as being stable and non-volatile.

[0207] These advantages are retained for other implementations of neuron 28, such as the one shown in Figure 6.

[0208] In such a case, instead of being parallel, the tracks 32 converge at one of their ends towards a detection zone 46 which forms a reservoir common to all the tracks 32.

[0209] Such a detection zone 46 then corresponds to a chamber where all the magnetic particles 40 accumulate.

[0210] This makes it possible to limit the magnetic tunnel junction surface and localize the detection so as not to impact the rest of the device, but is not compatible with an injection of magnetic particles 40 solely by displacement from a reservoir because the magnetic particles 40 of the different tracks 32 are then mixed in the chamber.

[0211] According to another embodiment, the neuron 28 further comprises a reinjection unit. The reinjection unit is capable of reinjecting the output value into neurons 28 located elsewhere in the neural network 28.

[0212] In each case, the described neuron 28 thus makes it possible to spatially condense the architecture with all neuromorphic functions, to reduce the need for subsequent electrical connections between neuromorphic components 28 and to directly integrate the artificial neurons 28 and synapses 30 into a microscopic architecture.

[0213] In fact, with the neuromorphic circuit 10, all of the operations necessary for neuromorphic computing are performed by the same device, which allows a large saving of surface area on the substrate.

[0214] Moreover, since skyrmions are 40 submicron particles, the dimensions of the integrated neuromorphic circuit 10 could be small enough to be integrated on a chip.

[0215] The present neuromorphic circuit 10 also makes it possible to ensure good compatibility between the output of the synapses 30 and the input of the neurons 28.

[0216] The neuromorphic component physically realizing the neuron 28 also makes it possible to reduce the number of operations carried out by directly adding up the number of local magnetic modifications in the detection zone 46.

[0217] Furthermore, for the configuration with magnetic tunnel junctions, the number of magnetic tunnel junctions is much lower than devices involving the presence of one magnetic tunnel junction per synapse 30.

[0218] Furthermore, the neuromorphic circuit 10 allows good isolation of the synapses 30 by preventing in particular currents corresponding to a synaptic weight from propagating into another synapse 30 by passing through the detection zone 46.

[0219] Thus, the neuromorphic circuit 10 provides fully integrated, compact artificial synapses 30 and neurons 28 that are both electrically controllable and detectable. This makes it possible to envisage the realization of a complete neuromorphic circuit 10 on a single chip.

[0220] For example, with reference to figure 7, the neuromorphic circuit 10 may comprise m neurons 28 connected by n synapses 30.

[0221] In this schematic representation, neurons 28 and synapses 30 are directly connected to each other, at the microscopic scale along continuous lines forming tracks 32.

[0222] Here the upper end (opposite the end where the current pulses are injected) of the tracks 32 is connected to ground. The resistance of the tracks 32 being negligible compared to that of the tunnel barrier, the measurement of the magnetic tunnel junction can be done between the upper electrode (oval in Figure 7) and ground, while avoiding short-circuiting the tracks 32 between them.

[0223] Alternatively or additionally, it is possible to arrange different layers of neurons 28 in space. This can be done for example in the X and Y plane, by recovering the m output voltages from Figure 7 to generate m input currents for the next layer arranged perpendicularly.

[0224] It is also possible to superimpose the layers of neurons 28 and synapses 30 in the stacking direction Z with the successive layers rotated by an angle of 90° between them.

[0225] Such a neuromorphic circuit 10 is thus capable of performing all neuromorphic operations (computation and learning) locally while being completely electrically controlled and consuming a limited amount of energy, compared to known neuromorphic circuits. Such neuromorphic circuits 10 can then be used for sensory recognition applications (sound, visual, touch, etc.) or to develop artificial intelligence systems capable of learning, reasoning or interacting.

[0226] Those skilled in the art will understand that the invention is not limited to the examples described in the description and that the embodiments described above are likely to be combined when such a combination is technically possible.

Claims

CLAIMS 1. Neuromorphic component (28) physically realizing a neuron of a neural network connected to at least one synapse (30), each synapse (30) carrying a respective weight, the neuron (28) taking as input at least one input value weighted by a respective weight to obtain an output value, the neuromorphic component (28) comprising: - for each input value: - a track (32) capable of presenting local magnetic modifications (40) which can be counted and moved, - a control unit (34) for controlling the local magnetic modifications (40), the control unit (34) being capable of annihilating and creating the local magnetic modifications (40), the control unit (34) also being capable of bringing the local magnetic modifications (40) into a detection zone (46) for the local magnetic modifications (40), the detection zone (46) being common to all the tracks (32), the control unit (34) comprising a sub-unit (42) for applying electrical pulses, the application sub-unit (42) being capable of applying a current proportional to the input value, and a sub-unit (44) for modifying at least one magnetic property of the track (32) with an amplitude proportional to the weight weighting the input value, and - a detection unit (36) for detecting local magnetic modifications (40), the detection unit (36) being capable of applying an activation function to an electrical quantity corresponding to the number of local magnetic modifications detected in the detection zone (46) to obtain the output value, the detection unit (36) being positioned on the detection zone (46).

2. A neuromorphic component according to claim 1, wherein the detection unit (36) uses a physical effect to obtain the activation function, the effect being a magnetotransport effect.

3. Neuromorphic component according to claim 2, wherein the detection unit (36) is a magnetic junction.

4. Neuromorphic component according to any one of claims 1 to 3, in which the tracks (32) are separate.

5. Neuromorphic component according to any one of claims 1 to 3, in which the detection zone (46) is a reservoir common to all the tracks (32).

6. Neuromorphic component according to any one of claims 1 to 5, in which the neuromorphic component (28) further comprises a reinjection unit, the reinjection unit being capable of reinjecting the output value into other neurons of the neural network.

7. Neuromorphic component according to any one of claims 1 to 6, in which the local magnetic modifications (40) are chosen from the list consisting of: - skyrmions, - hopfions, - bubbles, - blood cells, - cocoons, - strands, and - domain walls.

8. Neuromorphic component according to any one of claims 1 to 7, in which the detection zone (46) is made of a material being: - an element selected from Ni, Fe, Co, Ge, B, C, P, S, Mn, Cr, I, Br, Si, Pt, Al, Ga, Ta, Ru, Ht, Mg, Bi, Mo, Te, Se, W, N, Ti, Ir, V, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Er, Tm, and Yb, - an alloy comprising one or more of the aforementioned elements, and - an oxide comprising one or more of the above-mentioned elements.

9. Neuromorphic component according to any one of claims 1 to 8, in which each track (32) is made of a material chosen from: - an element chosen from 10. Ni, Fe, Co, Ge, B, C, P, S, Mn, Cr, I, Br, Si, Pt, Al, Ga, Ta, Ru, Ht, Mg, Bi, Mo, Te, Se, W, N, Ti, Ir, V, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Er, Tm and Yb - an alloy comprising one or more of the aforementioned elements, and - an oxide comprising one or more of the above-mentioned elements.

10. A neuromorphic component according to any one of claims 1 to 9, wherein the modification subunit (44) uses a physical effect selected from an electrostatic effect, a magneto-ionic effect, an electro-ionic effect, an electrochemical effect, a piezoelectric effect and a magnetostrictive effect.

11. Neuromorphic circuit (10) physically realizing a neural network, the neuromorphic circuit (10) comprising a set of neuromorphic components (28) according to any one of claims 1 to 10.

12. A neuromorphic circuit according to claim 11, wherein the set of neuromorphic components (28) is arranged to form a deep network.

Images