Apparatus and method for signal transduction for artificial neuron applications
By designing magnetic components and bonding members, the initial signal is converted into a dataset by utilizing the movement of magnetic domain walls, thus solving the problems of large size and high energy consumption of neuromorphic network devices and achieving compact and efficient signal conversion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GOLANA COMPUTING CORP
- Filing Date
- 2024-12-16
- Publication Date
- 2026-08-04
AI Technical Summary
Existing technologies for realizing complete neuromorphic networks result in large device sizes and high energy consumption, making it difficult to effectively process multiple signals.
A conversion device comprising magnetic components, bonding members, an initialization unit, an excitation unit, and a detection unit is employed. The initial signal is converted into a dataset by the movement of the magnetic domain walls, and a compact and energy-efficient signal conversion is achieved using two chambers.
A compact neuromorphic network conversion device was achieved, reducing energy consumption while maintaining good signal conversion quality.
Smart Images
Figure CN122514767A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a classification system designed to categorize input signals into signal classes. This type of classification system is particularly well-suited for applications in the development of spiking neural networks and is known as "in-material computing." This refers to the use of physical devices to artificially reproduce the behavior of neural networks, where the properties of the physical devices under consideration are utilized, particularly the properties of the materials they comprise.
[0002] The present invention also relates to a classification method for classifying input signals into signal categories, the classification method being implemented by the classification system as described above. Background Technology
[0003] There are different ways to artificially reproduce neural networks, for example, in the field of photonics or in the field of spintronics.
[0004] Existing solutions in the field of photonics have enabled the replication of the behavior of individual components as well as the demonstration of the functionality of more complete devices. However, these implementations have complex architectures and typically involve high energy consumption.
[0005] Existing solutions in spintronics are also of interest because they exhibit good durability relative to photonic systems and are compatible with CMOS (complementary metal-oxide-semiconductor) technology.
[0006] However, to date, the various solutions have been limited to proof-of-concept demonstrations of simple devices and are not suitable for the realization of functional neuromorphic devices.
[0007] Document EP4160484A1 discloses a functional neuromorphic network specifically capable of performing speech recognition of dictated digits. The network consists of a set of magnetic devices, each reproducing the behavior of a neuron. Specifically, these magnetic devices are configured to reproduce the properties of accumulation, loss, or leakage, as well as triggering or “ignition” exceeding a certain threshold.
[0008] To reconstruct the complete neuromorphic network, chambers are arranged in multiple rows and separated by gates, with all rows connected to a central nucleation chamber that serves as the starting point for the propagation of domain walls. This propagation of domain walls then occurs when a physical excitation corresponding to the signal to be processed is generated. To achieve this, magnetic domains of a given orientation are nucleated in the central nucleation chamber. This magnetic orientation is opposite to that of the rest of the magnetic components, thereby enabling the formation of domain walls. In this configuration, a physical excitation will cause the magnetic domain walls located at the interface between two domains with opposite magnetization to be moved. Thus, chambers are filled and gates are crossed until a network is obtained in which some chambers are in a first magnetic orientation and others are in a second magnetic orientation. The resulting image of the complete network then possesses characteristics of the information to be processed for stimulating the system.
[0009] This solution has the advantage of allowing signal processing without requiring very large computing power, especially for preprocessing calculations and for calculating synaptic weights.
[0010] However, to obtain neuromorphic networks capable of distinguishing a wide variety of signals, it is necessary to increase the number of rows in the chambers. However, given that the size of the chambers ranges from 0.1 µm to 10 µm and the number of rows can exceed 10,000, the total surface area of the classification system may rapidly increase, making it unsuitable for some applications.
[0011] Furthermore, the energy consumption required for physical excitation by a magnetic field increases with the surface area of the device. When an electric current is used as the physical quantity for excitation, the energy consumption increases with the length of the row.
[0012] Therefore, there is a need to find a compact and energy-efficient solution that can reproduce complete neuromorphic networks. Summary of the Invention
[0013] The purpose of this invention is to provide a solution to all or part of the aforementioned problems.
[0014] This objective can be achieved by an embodiment of a conversion device for converting an initial signal into a dataset, the conversion device comprising:
[0015] - A magnetic component, comprising a magnetic material configured to locally vary between a first magnetic state having a first magnetization and a second magnetic state having a second magnetization different from the first magnetization, the magnetic component being subdivided into a plurality of regions, wherein each region is in either the first magnetic state or the second magnetic state; each region in the first magnetic state is separated from the region in the second magnetic state by magnetic domain walls;
[0016] - A bonding member that separates the magnetic components between the upstream and downstream chambers and ensures magnetic communication between the upstream and downstream chambers. The bonding member is characterized by a magnetic transfer value that corresponds to the ability of the bonding member to allow displacement of the domain walls in the propagation direction, which is defined as from the upstream chamber toward the downstream chamber.
[0017] - An initialization unit configured to place the magnetic components in an initial magnetic configuration in which domain walls are provided at the bonding member;
[0018] - An excitation unit configured to generate a physical excitation to excite the magnetic material, thereby moving the domain walls in the propagation direction, the physical excitation being generated based on an initial signal;
[0019] - A detection unit, disposed in the downstream chamber, is configured to detect and record at least one feature triggering parameter when a physical parameter associated with the deformation state of the domain wall exceeds a predetermined threshold, the dataset including the at least one feature triggering parameter.
[0020] The setup described above enables the provision of a conversion device capable of transforming an initial signal into a dataset including one or more feature-triggered parameters. This type of device is particularly well-suited for artificial neuron applications. Furthermore, the use of only two chambers allows for a more compact conversion device requiring lower electrical input.
[0021] The term "magnetic transfer value" means the ability of a joint member to transmit magnetic domain walls. In other words, "magnetic transfer value" corresponds to the reciprocal of the resistance that a joint member can generate when passing through a magnetic domain wall.
[0022] The term "magnetic domain wall" refers to the interface that separates two regions in the same continuous magnetic medium that are not in the same magnetic state.
[0023] The conversion device may also have one or more of the following features, either individually or in combination.
[0024] According to one embodiment, the conversion device includes a nucleating member configured to place a magnetic component in a magnetic nucleation configuration in which all regions of the magnetic component are placed in a first magnetic state, except for one nucleating region placed in a second magnetic state, the nucleating region being contained in an upstream chamber.
[0025] Typically, the magnetic material of a magnetic component is heterogeneous. In other words, in most configurations, the magnetic component comprises multiple regions, wherein at least one of these regions is in a first magnetization state, and wherein at least one other region of these regions is in a second magnetization state.
[0026] According to one embodiment, the initialization unit includes a nucleating component.
[0027] According to one embodiment, the physical parameter associated with the deformation state of the domain wall is the distance measured from the bonding member in the propagation direction. When this distance exceeds a predetermined threshold distance, the feature triggering parameter is detected and recorded by the detection unit.
[0028] Alternatively, the physical parameter associated with the deformation state of the domain wall is a measurement of the surface area or surface area ratio, and when the surface area or surface area ratio exceeds a predetermined threshold, the feature triggering parameter is detected and recorded by the detection unit.
[0029] According to one embodiment, the dataset is the number corresponding to the number of times the feature triggering parameters are recorded by the detection unit. In other words, the dataset is equal to the sum of the detected feature triggering parameters.
[0030] According to one embodiment, the excitation unit is configured to at least temporarily stop the physical excitation of the magnetic material when the following condition is met: at least one characteristic triggering parameter is detected by the detection unit.
[0031] The ability to select a threshold exceeding which the characteristic triggering parameter is detected allows for detection of this parameter even when the domain walls have not fully filled the downstream chamber. Therefore, the system can passively return to its initial configuration when the excitation unit temporarily stops the physical excitation of the magnetic material. In this way, a series of chambers can be artificially reproduced.
[0032] According to one embodiment, the conversion device further includes a closed contour outer perimeter boundary, within which the magnetic component is completely contained.
[0033] Therefore, when a domain wall encounters an outer boundary, its expansion can be restricted.
[0034] Typically, the outer perimeter is a non-magnetic region. For example, it could be a trench etched into the material constituting a magnetic component, or a region that has been irradiated by a laser beam or ions, thus enabling the formation of the non-magnetic region.
[0035] According to one embodiment, the outer perimeter is a quadrilateral. Therefore, the manufacturing of the conversion device is simplified.
[0036] According to one embodiment, the initialization unit is configured to place the magnetic component in the initial magnetic configuration when the following condition is met: at least one feature triggering parameter is detected by the detection unit.
[0037] The ability to reset the magnetic component by placing it back into the initial magnetic configuration after each detection of a characteristic trigger parameter allows for a faster return of the magnetic component to that initial configuration, enabling the conversion device to detect a maximum number of characteristic trigger parameters in time. Furthermore, this allows for the manual reproduction of a conversion device comprising a series of chambers.
[0038] According to one embodiment, the initialization unit includes an extension member configured to move a magnetic domain wall until it engages a member to place the magnetic component in an initial magnetic configuration.
[0039] Therefore, the conversion device can be placed in a configuration in which the physical excitation generated by the excitation unit only helps to move the magnetic domain walls across the connecting member into the downstream chamber.
[0040] According to one embodiment, the initialization unit includes an excitation unit configured to move the domain wall until it engages the member, so as to place the magnetic component in an initial magnetic configuration.
[0041] According to one embodiment, the conversion device further includes a reconstruction unit configured to modify the magnetic transmission value of the coupling member.
[0042] Therefore, the magnetic transfer value of the connecting member can be modified to adapt it to the initial signal for the desired conversion.
[0043] According to one embodiment, the reconstruction unit is configured to modify the magnetic transmission value of the joint member whenever the detection unit detects a feature trigger parameter.
[0044] Therefore, when the magnetic components are placed in the initial magnetic configuration, the magnetic transfer value of the bonding member can be modified. This makes it possible to artificially recreate a series of chambers separated by bonding members with varying magnetic transfer values. In other words, the bonding member is reconfigurable.
[0045] According to one embodiment, the excitation unit includes a current generator configured to generate current.
[0046] In a coordinated manner, a current generator is used as the excitation unit in a converter with only two chambers, which enables the energy consumption of the system to be reduced while maintaining good conversion quality of the initial signal by the converter.
[0047] According to one embodiment, the physical excitation corresponds to a signal including a current pulse.
[0048] The object of the present invention can also be achieved by implementing a classification system designed to classify input signals into signal categories, the classification system comprising:
[0049] - At least one preprocessing device designed to transform an input signal into at least one initial signal;
[0050] - At least one conversion device as described above, which receives an initial signal from the at least one initial signal as input and converts the initial signal received as input into a dataset;
[0051] - Identification unit, which receives a dataset as input and is configured to associate an input signal with a signal category based on the dataset, the signal category being selected from a set of predetermined signal categories stored in the memory of the classification system.
[0052] The arrangement described above enables the provision of a classification system capable of classifying input signals based on responses detected by the conversion device.
[0053] According to one embodiment, the classification system includes at least two conversion devices.
[0054] According to one embodiment, the classification system includes at least two conversion devices that receive the same initial signal as input, or each receives a different initial signal as input obtained by transforming the same input signal using different preprocessing devices.
[0055] For example, the classification system can receive a single input signal as input and may include one or more preprocessing units. The preprocessing unit can convert the input signal into an initial signal that is the same for all conversion units, or alternatively, convert it into multiple different initial signals, which may be different or the same for each conversion unit. In other words, the same input signal can be preprocessed by the preprocessing unit in a different manner for each conversion unit.
[0056] In this way, at least two transformations from the initial signal to the dataset can be performed to improve the signal transformation.
[0057] The object of the present invention can also be achieved by implementing a method for converting an initial signal into a dataset, the conversion method being implemented by a conversion device as described above, and comprising the following stages:
[0058] - Initialization phase, including initial steps, in which the magnetic components are placed in an initial magnetic configuration;
[0059] - Excitation phase, in which the excitation unit generates physical excitation based on the initial signal in order to move the magnetic domain walls;
[0060] - Reconstruction phase, which is implemented when the detection unit detects that at least one physical parameter associated with the deformation state of the magnetic domain wall exceeds a predetermined threshold, includes a detection step in which the detection unit records at least one feature triggering parameter in the dataset.
[0061] The arrangement described above enables a conversion method for converting an initial signal into a dataset using a conversion device that can utilize only two chambers made of magnetic material.
[0062] Therefore, it is easy to understand that the excitation phase is implemented continuously, and multiple detection steps can be implemented, especially whenever at least one physical parameter associated with the deformation state of the domain wall exceeds a predetermined threshold.
[0063] The conversion method may also have one or more of the following features, either individually or in combination.
[0064] In one embodiment, the initial signal and physical stimulus are time-dependent. In this case, the conversion method can stop when the time for the physical stimulus ends.
[0065] According to one embodiment, the initialization unit is configured to place the magnetic components in an initial magnetic configuration, for example, by means of an extension member.
[0066] According to one embodiment, the initialization unit includes a nucleating member, and the initialization phase then includes a nucleation step performed prior to the initialization step, in which the nucleating member places a magnetic component in a magnetic nucleation configuration in which all regions of the magnetic component are placed in a first magnetic state, except for one nucleation region placed in a second magnetic state, the nucleation region being contained in an upstream chamber.
[0067] Therefore, the conversion device can be prepared to perform the conversion, especially when it has not been used for a long time.
[0068] According to one embodiment, the initial step is performed simultaneously with the excitation phase. In other words, it is the physical excitation generated during the excitation phase that enables the domain walls to move toward the bonding member in order to transition from a magnetic nucleation configuration to an initial magnetic configuration.
[0069] According to one embodiment, the initial magnetic configuration corresponds to a configuration in which the heterogeneous magnetic material is in a second magnetic state in the upstream chamber and in a first magnetic state in the downstream chamber, or vice versa.
[0070] According to one embodiment, the reconstruction phase is performed when the magnetic material present in the downstream chamber is fully in the second magnetic state.
[0071] Therefore, it is simpler to detect when the magnetic domain wall exceeds the threshold distance.
[0072] According to one embodiment, the reconfiguration phase includes a stop step in which the excitation unit at least temporarily suspends the implementation of the excitation phase; a new excitation phase is implemented at the end of the reconfiguration phase.
[0073] The stop excitation phase allows the magnetic components to return to their initial magnetic configuration before the excitation phase is repeated.
[0074] According to one embodiment, the new initialization phase is implemented at the end of the refactoring phase.
[0075] Therefore, even when the downstream chamber is fully in the second magnetic state, the magnetic components can be placed in the initial magnetic configuration. The magnetic components can be actively returned to the initial magnetic configuration to save time.
[0076] According to one embodiment, the reconstruction phase further includes a modification step implemented after the detection step, in which the reconstruction unit modifies the magnetic transmission value of the joint member.
[0077] In this way, a series of chambers separated by jointing members with varying magnetic transmission values can be artificially reproduced.
[0078] The object of the present invention can also be achieved by implementing a classification method for classifying input signals into signal categories, which is implemented using a classification system as described above and includes the following steps:
[0079] - A preprocessing step, in which the preprocessing device transforms the input signal into at least one initial signal;
[0080] - At least one conversion step, in which a conversion device applies the conversion method as described above to an initial signal among the at least one initial signal in order to obtain a dataset corresponding to the initial signal to which the conversion method is applied;
[0081] -Identification step, in which the identification unit associates the input signal with a signal category based on a dataset obtained during at least one transformation step.
[0082] The arrangement described above enables the provision of a method for classifying input signals into signal categories using a compact conversion device.
[0083] According to one embodiment, the preprocessing step may include transforming the input signal into a plurality of initial signals. Each of the plurality of initial signals may be the same as or different from another of the initial signals. Thus, during each conversion step, each conversion device applies a conversion method to its associated initial signal.
[0084] In this way, the input signal can be preprocessed in different ways during the preprocessing step, either according to the features to be extracted from the input signal or according to the conversion device that applies the conversion method to the initial signal thus preprocessed. Attached Figure Description
[0085] Other aspects, objects, advantages, and features of the invention will become more apparent from the following detailed description of preferred embodiments of the invention, which is given by way of non-limiting example and with reference to the accompanying drawings, in which:
[0086] [ Figure 1 ] Figure 1 This is a schematic diagram of a classification system according to a specific embodiment of the present invention.
[0087] [ Figure 2 ] Figure 2 This is a schematic diagram illustrating different ways of carrying out the nucleation step through nucleating components.
[0088] [ Figure 3 ] Figure 3 This is a schematic diagram of a conversion apparatus according to a specific embodiment of the present invention, which performs certain steps of a conversion method.
[0089] [ Figure 4 ] Figure 4 This is a schematic perspective view of a conversion device according to a specific embodiment of the present invention, and more particularly shows the detection unit.
[0090] [ Figure 5 ] Figure 5 It is a graph showing the evolution of physical parameters associated with the deformation state of the domain wall over time.
[0091] [ Figure 6 ] Figure 6 These are schematic side and top views of a conversion device according to a specific embodiment of the present invention, the conversion device including a fixed engagement member.
[0092] [ Figure 7 ] Figure 7 These are schematic side and top views of a conversion device according to a specific embodiment of the present invention, the conversion device including a reconfigurable engagement member.
[0093] [ Figure 8 ] Figure 8This is a schematic diagram illustrating different ways of implementing modification steps through a refactoring unit.
[0094] [ Figure 9 ] Figure 9 This is a schematic diagram illustrating different ways of implementing modification steps through a refactoring unit.
[0095] [ Figure 10 ] Figure 10 This is a schematic diagram illustrating some steps in the initialization phase.
[0096] [ Figure 11 ] Figure 11 This is a schematic diagram of a classification method according to a specific embodiment of the present invention.
[0097] [ Figure 12 ] Figure 1 This is a schematic diagram of a classification method according to a specific embodiment of the present invention. Detailed Implementation
[0098] In the remainder of the accompanying drawings and description, the same reference numerals denote the same or similar elements. Furthermore, different elements are not drawn to scale to ensure clarity of the drawings. Moreover, different embodiments and variations are not mutually exclusive and can be combined with each other.
[0099] like Figures 1 to 9 As shown, the present invention relates to a conversion device 10 for converting an initial signal (denoted as "SI") into a dataset (denoted as "SC"). For example, the initial signal SI may correspond to an input signal (denoted as "SE") that has been preprocessed by a preprocessing device 3, which will be described later. The input signal SE may correspond to information to be processed, which has been preprocessed into the initial signal SI that can be converted by the conversion device 10. For example, and not limited to, the input signal SE may be preprocessed into a series of pulses, as described in European Patent Application No. 22315278.6, filed in November 2022. Alternatively, the initial signal SI may correspond to variations in amplitude, frequency, time response, or any other type of signal that enables the excitation of magnetic materials.
[0100] The conversion device first includes a magnetic component 20 comprising a magnetic material configured to locally vary between a first magnetic state having a first magnetization (denoted as "M1") and a second magnetic state having a second magnetization different from the first magnetization (denoted as "M2"). Figure 1As can be seen, the conversion device 10 may include a closed contour outer perimeter boundary 11, within which the magnetic component 20 is completely contained. Therefore, and as will be presented later, the expansion of the domain wall P can be limited when it encounters the outer perimeter boundary 11. Typically, the outer perimeter boundary 11 is a non-magnetic region, for example, a trench etched into the material constituting the magnetic component 20, or a region that has been irradiated by a laser beam or ion irradiation to enable the formation of the non-magnetic region. It is also possible, but not limited to, that such an outer perimeter boundary 11 is quadrilateral. Therefore, the manufacture of the conversion device 10 is simplified.
[0101] The magnetic component 20 is subdivided into multiple regions, each of which is in either a first magnetic state M1 or a second magnetic state M2. Each region in the first magnetic state M1 is separated from a region in the second magnetic state M2 by a domain wall P. The term "domain wall P" means an interface that separates two regions in the same continuous magnetic medium that are not in the same magnetic state. Typically, the magnetic material of the magnetic component 20 is heterogeneous. In other words, in most configurations, the magnetic component 20 includes multiple regions, wherein at least one of these regions is in the first magnetic state M1, and wherein at least one other region of these regions is in the second magnetic state M2.
[0102] The conversion device 10 also includes a coupling member 30 that separates the magnetic component 20 between the upstream chamber 21 and the downstream chamber 23. Figure 1 In a particular embodiment, the upstream chamber 21 is fully in a second magnetic state M2, and the downstream chamber 23 is fully in a first magnetic state M1. A connecting member 30, characterized by a magnetic transfer value (denoted as "Rtr"), ensures magnetic communication between the upstream and downstream chambers 21 and 23. This magnetic transfer value corresponds to the ability of the connecting member 30 to allow displacement of the domain wall P in the propagation direction (denoted as "X"), defined as from the upstream chamber 21 to the downstream chamber 23. The term "magnetic transfer value Rtr" signifies the ability of the connecting member 30 to allow the transmission of the domain wall P. In other words, "magnetic transfer value Rtr" corresponds to the reciprocal of the resistance that the connecting member 30 can generate as the domain wall P passes through. Typically, the propagation of the domain wall P within one of the two chambers 21 or 23 is easier than its propagation through the connecting member 30. Therefore, the connecting member 30 acts as a gate, constituting a brake on the propagation of the domain wall P between the upstream and downstream chambers 21 and 23.
[0103] The conversion device 10 also includes an initialization unit 40 configured to place the magnetic component 20 in an initial magnetic configuration (referred to as "C1"), in which domain walls P are disposed at the bonding member 30. Such an initial magnetic configuration C1 is, for example, in... Figure 1 The Chinese side indicated that...
[0104] According to one embodiment, the conversion device 10 includes a nucleating member 41 configured to place the magnetic component 20 in a magnetic nucleation configuration (denoted as C0), in which all regions of the magnetic component 20 are placed in a first magnetic state M1, except for one nucleation region placed in a second magnetic state M2, which is contained within an upstream chamber 21. The initialization unit 40 may also include the nucleating member 41. Figure 2 Different variations of the nucleating member 41 are shown. Placing the magnetic component 20 in the magnetic nucleation configuration C0 is particularly useful when the entire conversion device 10 is in the first magnetic state M1. In this case, the nucleating member 41 enables the formation of regions in the second magnetic state M2, thereby creating magnetic domain walls P.
[0105] Figure 2 A shows a nucleating member 41 configured to apply a local magnetic field to form a nucleation region. Typically, the nucleating member 41 may include an electromagnet or permanent magnet device that enables the application of a local magnetic field. Figure 2 B illustrates a nucleation member 41 configured to apply a local magnetic field associated with the presence of intrinsic or intentionally created defects 42 in the magnetic material. The creation of defects 42 in the upstream chamber 21 thus serves as a preferred center for promoting nucleation. Specifically, the defects 42 can be created by locally modifying the properties of the magnetic layer using various means, such as laser beams, FIB beams (for focused ion beams), or optical or electron beam lithography steps followed by etching steps. Figure 2 C illustrates a nucleating member 41 configured to apply a localized magnetic field in association with the presence of a magnetic layer 44 attached to the magnetic component 20. A non-magnetic layer 46 may or may not be inserted between the magnetic component 20 and the magnetic layer 44. The magnetic layer 44, or the stack of layers 44 and 46, may have planar magnetic anisotropy. Figure 2 D illustrates a nucleating member 41 configured to apply a local or global magnetic field associated with local heating. The nucleating member 41 may then include a local heating device 47 configured to generate such local heating by means of an electric current, a localized laser beam, or any other means. Figure 2 E shows a nucleating member 41 configured to apply a local magnetic field in the presence of a local electric field. In this case, the electric field can be used to lower the magnetic anisotropy barrier. Finally, Figure 2 F shows a nucleating member 41, which is configured to inject spin-polarized current, for example, by defining a column of magnetically stacked magnetic tunnel junctions or spin valves on the upstream chamber 21. Then, as Figure 2 The injected current is indicated by F. It is understood that these different embodiments of the nucleating member 41 are not exhaustive.
[0106] According to one embodiment, the initialization unit 40 includes an extension member 43 configured to move the domain wall P to the engagement member 30 to place the magnetic component 20 in the initial magnetic configuration C1. Thus, the conversion device 10 can be positioned in a configuration where the physical excitation generated by the excitation unit 50, described below, only facilitates the movement of the domain wall P across the engagement member 30 into the downstream chamber 23. According to a first variation, the extension member 43 is configured to move the domain wall P in the propagation direction X, for example, once the nucleating member 41 has formed a nucleation region. According to another variation, the extension member 43 can be configured to move the domain wall P in the opposite direction to the propagation direction X, for example, returning it to the engagement member 30 after the domain wall P has moved into the downstream chamber 23. Therefore, the arrangement described above allows the extension member 43 to reposition the magnetic component 20 in the initial magnetic configuration C1.
[0107] Different variations can be used to implement the extension member 43. According to a first variation, the extension member 43 includes a magnetic field generator, such as an electromagnet or permanent magnet device, enabling the application of a local magnetic field. According to a second variation, the extension member 43 includes a spin-polarized current generator. The extension member 43 may also include a nucleating member 41.
[0108] The conversion device 10 also includes an excitation unit 50 configured to generate a physical excitation to excite the magnetic material, thereby moving the domain wall P in the propagation direction X. This physical excitation is generated based on an initial signal SI. For example, the physical excitation corresponds to a signal including a current pulse or a magnetic field pulse. For this purpose, the excitation unit 50 may include a current generator 51 configured to generate a current, or a magnetic field generator.
[0109] According to one embodiment, the initialization unit 40 includes an excitation unit 50, which is then configured to move the domain wall P until it reaches the engagement member 30 to place the magnetic component 20 in the initial magnetic configuration C1. In other words, the excitation unit 50 includes an extension member 43.
[0110] According to one embodiment, the excitation unit 50 can be configured to place the magnetic component 20 in the initial magnetic configuration C1 when the magnetic domain wall P has passed the engagement member 30 and is positioned in the downstream chamber 23.
[0111] In a coordinated manner, the use of a current generator 51 as an excitation unit 50 in the conversion device 10, which comprises only two chambers 21 and 23, reduces the energy consumption of the system while maintaining good conversion quality of the initial signal SI by the conversion device 10. Furthermore, the current generator 51 can function as both an extension member 43 and an excitation unit 50. Figure 3 The conversion device 10 is shown according to different configurations. Figure 3 A shows the conversion device 10 when the magnetic component 20 is in the nucleation configuration C0. Figure 3 B illustrates the conversion device 10 when the magnetic component 20 is in the initial configuration C1. The current generator 51 can be configured as follows:
[0112] - Inject current to move the domain wall P in the propagation direction X so as to switch the magnetic component 20 from the nucleation magnetic configuration C0 to the initial magnetic configuration C1;
[0113] - Inject current to move the domain wall P across the bonding member 30 in the propagation direction X;
[0114] - Inject a reverse current to move the domain wall P in the opposite direction to the propagation direction X, so as to return the domain wall P to the joining member 30 in the initial magnetic configuration C1.
[0115] Therefore, these three functions are implemented by a single component.
[0116] The conversion device 10 also includes a detection unit 60 disposed in the downstream chamber 23. The detection unit 60 is configured to detect and record at least one characteristic trigger parameter (denoted as "Nbf") when a physical parameter associated with the deformation state of the domain wall P exceeds a predetermined threshold (denoted as "Vs"). Different variations can be used to implement the detection unit 60. According to a first variation, the detection unit 60 includes an optical detector, such as a wide-field or focused Kerr effect optical microscope. Figure 4 A second variation is shown, in which the detection unit 60 operates via electrical detection. In this case, the detection unit 60 can electrically detect at least one trigger parameter by means of a magnetic tunnel junction (tunneling magnetoresistance signal) or a spin valve (giant magnetoresistance signal). More specifically, the detection unit 60 may include a pillar 61 arranged on the magnetic component 20 directly above the downstream chamber 23. The pillar 61 need not be centered relative to the downstream chamber 23. In fact, and advantageously, the position of the pillar 61 relative to the engagement member 30 helps to define a predetermined threshold VS. According to one embodiment, the pillar 61 may include a spin valve-type stack or a magnetic tunnel junction-type stack, which includes:
[0117] - Contact electrode 62 used for reading electrical signals;
[0118] - The insulating material layer 63 (e.g., MgO, AlOx, HfO2, etc. in the case of magnetic tunnel junction stacking) or the non-ferromagnetic conductive material layer in the case of spin valve type stacking;
[0119] - A magnetically stacked layer 64, which constitutes a reference magnetic electrode.
[0120] According to the first variant, the contact electrode 62 is a magnetic layer magnetically coupled to the magnetic material constituting the magnetic component 20, and the local magnetic state is then read at the spin valve-type stack or the magnetic tunnel junction-type stack. Alternatively, the magnetic layer 62 can be omitted. In the case of a magnetic tunnel junction-type stack, an insulating material layer 63, or in the case of a spin valve-type stack, a non-ferromagnetic conductive material layer 63, is deposited directly on the magnetic material constituting the magnetic component 20. The local magnetic state is then read at the spin valve-type stack or the magnetic tunnel junction-type stack.
[0121] The detection unit 60 may also include an electrical device that allows reading the resistance of the cylinder 61, and Figure 4 Contacts not shown in the diagram.
[0122] According to one embodiment, the physical parameter associated with the deformation state of the domain wall P is the distance (denoted as "d") measured from the joint member 30 in the propagation direction X. When the distance d exceeds a predetermined threshold distance VS, the feature triggering parameter Nbf is detected and recorded by the detection unit 60. Figure 3 C to Figure 3 F illustrates the switching device 10 as the domain wall P moves within the downstream chamber 23. According to one possibility, the predetermined threshold VS is a fixed distance measured between the position of the engaging member 20 and its location within the downstream chamber 23, such as... Figure 3 C and Figure 3 As shown in Figure D. In this case, when the domain wall P exceeds a fixed distance VS, the triggering parameter Nbf can be detected, as shown in Figure D. Figure 3 As shown in D. Because in Figure 3 Since C did not exceed the fixed distance VS, the feature triggering parameter Nbf was not detected.
[0123] Alternatively, the physical parameter associated with the deformation state of the domain wall P can be a surface area measurement or a surface area ratio. When the surface area or surface area ratio exceeds a predetermined threshold VS, the characteristic trigger parameter Nbf is then detected and recorded by the detection unit 60. For example, the predetermined threshold VS can correspond to the maximum area Amax of the downstream chamber 23, and the physical parameter can correspond to the area A of the region in the downstream chamber 23 that is in the second magnetic state M2, such as... Figure 3As shown in D. In this case, when the entire downstream chamber 23 is in the second magnetic state M2, that is, when A = Amax, or when the ratio A / Amax = 1, as shown in D. Figure 3 As shown in E, the feature triggering parameter Nbf can then be detected.
[0124] Figure 5 It is a graph showing the changes in physical parameters of the magnetic component 20 in relation to the deformation state of the domain walls in the downstream chamber 23, based on the current pulses generated by the physical excitation of the excitation unit 50.
[0125] Regardless of the variant considered, the dataset SC includes the at least one feature triggering parameter Nbf. For example, the dataset SC is the number corresponding to the number of times the feature triggering parameter Nbf is recorded by the detection unit 60. In other words, the dataset SC is equal to the sum of the detected feature triggering parameters Nbf.
[0126] Typically, the detection unit 60 may include a memory configured to store the dataset SC.
[0127] According to a non-limiting variation, the excitation unit 50 can be configured to at least temporarily stop the physical excitation of the magnetic material when at least one characteristic triggering parameter Nbf is detected by the detection unit 60. However, it is typically set so that the physical excitation is stopped by the excitation unit 50 only at the end of the time of the initial signal SI.
[0128] Advantageously, the initialization unit 40 can be configured to place the magnetic component 20 in the initial magnetic configuration C1 when at least one characteristic trigger parameter Nbf is detected by the detection unit 60. The magnetic component 20 can be reset by placing it in the initial magnetic configuration C1 after each detection of the characteristic trigger parameter Nbf, thus accelerating the return of the magnetic component 20 to the initial magnetic configuration C1 and enabling the conversion device 10 to detect a maximum number of characteristic trigger parameters Nbf in time. Furthermore, this allows for the manual reproduction of the conversion device 10, which includes a series of chambers. Figure 3 F shows a conversion device 10 in such a configuration that the initialization unit 40 is in the process of returning the magnetic component 20 to the initial magnetic configuration C1.
[0129] Although the magnetic transmission value Rtr of the joining member 30 can be fixed (e.g. Figure 6 As shown), but the magnetic transfer value Rtr can also be modified (as shown). Figure 7 As shown). In this case, and as... Figure 7As shown, the conversion device 10 may include a reconstruction unit 70 configured to modify the magnetic transfer value Rtr of the bonding member 30. This allows the bonding member 30 to be adapted to the initial signal SI to be converted. For example, the reconstruction unit 70 may be configured to modify the magnetic transfer value Rtr of the bonding member 30 whenever the detection unit 60 detects the characteristic trigger parameter Nbf. Therefore, the magnetic transfer value Rtr of the bonding member 30 can be modified when the magnetic component 20 is placed in the initial magnetic configuration C1. This allows for the artificial reproduction of a series of chambers separated by the bonding member 30 with varying magnetic transfer values Rtr. In other words, the bonding member 30 is reconfigurable. Different variations exist for reconfiguring the bonding member 30. Typically, the formation of the bonding member 30 or its reconstruction is implemented by controlling and modifying the local magnetic properties of the magnetic material constituting the magnetic component 20. The different variations presented below are different embodiments that enable those skilled in the art to implement the reconstruction unit and are not limiting.
[0130] Figure 8 A to Figure 8 B shows a reconstruction unit 70, which is capable of modifying the magnetic transfer value Rtr of the coupling member 30 by applying a local magnetic field. Figure 8 A more specifically shows a reconstruction unit 70 configured to apply a local magnetic field by means of a magnetic element 71 whose magnetization is perpendicular to the plane. The magnetic element 71 can be separated from the coupling member 30 by a non-magnetic element 73. Figure 8 Figure B illustrates a reconstruction unit 70 configured to apply a local magnetic field by means of a magnetic element 75 whose magnetization is located in a plane. The magnetic element 75 can be separated from the bonding member 30 by a non-magnetic element 73. Other variations, not shown, such as using a spin valve structure or a magnetic tunnel junction, can also be used to obtain the reconstruction unit 70.
[0131] Those skilled in the art may refer, for example, to European Patent Application No. 23315104.2 filed on April 25, 2023, to obtain the reconfiguration unit 70.
[0132] Figure 9 A shows a reconstruction unit 70 that can modify the magnetic transport value Rtr of the bonding member 30 by means of a local electric field generated, for example, by applying a voltage V. More specifically, this local electric field enables the modulation of the magnetic properties of the magnetic layer, particularly its anisotropy and Dzyaloshinskii-Moriya interaction. Figure 9 A prominently shows a conductive electrode 72, to which a voltage V is applied by means of a voltage generator 74. According to this embodiment, a dielectric layer 76 can be inserted between the conductive electrode 72 and the bonding member 30.
[0133] Figure 9 B and Figure 9 C illustrates a reconstruction unit 70, which is capable of modifying the magnetic transfer value Rtr of the bonding member 30 by means of localized heating. More specifically, Figure 9 B illustrates a conductive layer 77a into which a current (denoted as "I") is injected via a current generator 77b to raise the temperature of the conductive layer 77a through the Joule effect. The generated temperature is thus transmitted to the bonding member 30 directly or via an intermediate layer 78 disposed between the bonding member and the conductive layer 77a. The intermediate layer 78 may be, for example, a layer made of a dielectric material, to electrically insulate the bonding member 30 from the conductive layer 77a into which the current is injected. Localized heating may also be achieved by means of a hot tip (e.g., an atomic force microscope tip or AFM tip) near the bonding member, or by means of... Figure 9 This is achieved using a locally focused laser 77c, as shown in Figure C. For this purpose, an absorption layer 77d can be provided, configured to absorb the energy received from the laser 77c.
[0134] All the arrangements described above enable the provision of a conversion device 10 capable of converting an initial signal SI into a dataset SC including one or more feature triggering parameters Nbf. This type of device is particularly suitable for applications involving artificial neurons. Furthermore, the use of only two chambers 21, 23 allows for a more compact conversion device 10 that requires a lower power supply.
[0135] As previously indicated, the present invention also relates to a classification system 1 designed to classify an input signal SE into a signal category (denoted as "Ci"). Examples of such a classification system 1 are found in... Figure 1 As shown in the image.
[0136] Classification system 1 includes:
[0137] - At least one preprocessing device 3, which is designed to transform the input signal SE into an initial signal SI;
[0138] - At least one conversion device 10 as described above, which receives the initial signal SI as input and converts the initial signal SI into a dataset SC;
[0139] - Identification unit 5, which receives dataset SC as input and is configured to associate input signal SE with signal category Ci based on dataset SC, the signal category Ci being selected from a set of predetermined signal categories stored in memory 7 of classification system 1.
[0140] The arrangement described above enables the provision of a classification system 1 that can classify the input signal SE based on the response detected by the conversion device 10.
[0141] According to a variant not shown, the classification system 1 may include at least two transformation devices 10. These different transformation devices 10 may receive the same initial signal SI as input. Alternatively, the transformation devices 10 may each receive independent and different initial signals SI, which are obtained by transforming the same input signal SE by means of different preprocessing devices 3 associated with each of the transformation devices 10. In this way, the quality of classifying the input signal SE into the dataset SC can be improved. At least two transformations from the initial signal SI to the dataset SC may also be performed to improve signal transformation.
[0142] According to a first possibility, the classification system 1 may include a preprocessing unit 3 capable of converting the input signal SE into multiple different or identical initial signals SI. Each of the multiple initial signals SI is then associated with one of the transformation units 10. Thus, a different dataset SC can be obtained from each of the transformation units 10. According to a second possibility, the classification system 1 may include multiple preprocessing units 3, each designed to convert the input signal SE into an initial signal SI. Therefore, it is clearly understood that, for each of the transformation units 10, a different initial signal SI can be obtained from a single input signal SE.
[0143] The present invention also relates to a method for converting an initial signal SI into a dataset SC. Typically, the initial signal SI and the physical excitation are time-dependent. In this case, the conversion method can be configured to stop at the end of the physical excitation's time. This conversion method is implemented by a conversion device 10 as described above. Figure 3 as well as Figures 10 to 12 The different stages and steps of the conversion method are shown.
[0144] The conversion method includes an initialization phase P1.
[0145] According to one embodiment, the initialization unit 40 of the conversion device 10 may include a nucleating member 41. In this case, the initialization stage P1 may include a nucleation step E11, in which the nucleating member 41 places the magnetic component 20 in a nucleation magnetic configuration C0, in which all regions of the magnetic component 20 are placed in a first magnetic state M1, except for the nucleation region placed in a second magnetic state M2, which is contained in the upstream chamber 21. Therefore, the conversion device 10 can be prepared to perform conversion, especially when it has not been used for a long time.
[0146] The initialization phase P1 also includes an initial step E13, in which the magnetic component 20 is placed in an initial magnetic configuration C1. When the nucleation step E11 is performed therein, the initialization unit 40 may be configured, for example, to place the magnetic component 20 in the initial magnetic configuration C1 via an extension member 43. For example, the initial magnetic configuration C1 corresponds to a configuration in which the heterogeneous magnetic material is in a second magnetic state M2 in the upstream chamber 21 and a first magnetic state M1 in the downstream chamber 23, or vice versa.
[0147] The conversion method also includes an excitation phase P2, in which the excitation unit 50 generates a physical excitation based on the initial signal SI to move the magnetic domain wall P. Figure 10 As shown, the initial step E13 can be performed simultaneously with the excitation phase P2. In other words, it is the physical excitation generated during the excitation phase P2 that enables the magnetic domain wall P to move toward the bonding member 30 in order to switch from the nucleation magnetic configuration C0 to the initial magnetic configuration C1.
[0148] The conversion method also includes a reconstruction phase P3, which is implemented when the detection unit 60 detects that at least one physical parameter associated with the deformation state of the domain wall P exceeds a predetermined threshold Vs. The reconstruction phase P3 includes a detection step E31, in which the detection unit 60 records at least one feature triggering parameter Nbf in the dataset SC. Therefore, it is clear that the excitation phase P2 is implemented continuously, and multiple detection steps E31 can be implemented, particularly whenever at least one physical parameter associated with the deformation state of the domain wall P exceeds the predetermined threshold Vs. According to the first variation, the detection step can be implemented when the magnetic material present in the downstream chamber 23 is fully in the second magnetic state M2. This simplifies the detection of physical parameters associated with the deformation of the domain wall exceeding the predetermined threshold Vs. However, such a variation is not limiting, and the detection step E31 can also be implemented when the domain wall P crosses a threshold distance Vs, as... Figure 3 As shown in D.
[0149] According to a variant not shown, the reconfiguration phase P3 may include a stop step E32, in which the excitation unit 50 at least temporarily stops the implementation of the excitation phase P2; a new excitation phase P2 is implemented at the end of the reconfiguration phase P3. Stopping the excitation phase P2 allows the magnetic component 20 to return to the initial magnetic configuration C1 before reimplementing the excitation phase P2.
[0150] Whenever a characteristic trigger parameter is detected during detection step E31, it is advantageous to implement a new initialization phase P1 at the end of reconstruction phase P3. More specifically, once detection step E31 has been implemented, initialization unit 40 can return magnetic component 20 to initial magnetic configuration C1 by implementing a new initialization step E13. Therefore, it is clear that, according to this embodiment, a new initialization step is implemented whenever a detection step is implemented. As specified by the reference conversion device, when initialization unit 40 includes excitation unit 50, initialization step E13 can be implemented by excitation unit 50. In the specific case where excitation unit 50 includes current generator 51, it is advantageous that initialization step E13 implemented after detection step E31 is performed by injecting current from current generator 51 in the direction opposite to the propagation direction X. Therefore, magnetic component 20 can be actively returned to initial magnetic configuration C1 to save time. It is also clear that the implementation of nucleation step E11 is independent of the implementation of initialization step E13. For example, the conversion method may include a single nucleation step E11, and multiple initial steps E13 implemented (actively or otherwise) after each detection step E31.
[0151] According to a non-limiting variation, in this variation, the conversion device 10 includes a reconstruction unit, and the reconstruction stage P3 may further include a modification step E33 performed after the detection step E31, in which the reconstruction unit 70 modifies the magnetic transfer value Rtr of the coupling member 30. In this way, a series of chambers separated by coupling members having varying magnetic transfer values Rtr can be artificially reproduced. Such a variation is exemplified in… Figure 12 As shown in the image.
[0152] The arrangement described above enables a conversion method for converting an initial signal SI into a dataset SC by means of a conversion device 10, which can use only two chambers made of magnetic material.
[0153] Finally, and as Figure 11 and Figure 12 As shown, the present invention relates to a classification method for classifying an input signal SE into signal category Ci. This classification method is implemented by a classification system 1 as described above and includes the following steps:
[0154] - Preprocessing step E0, in which the preprocessing device 3 transforms the input signal SE into the initial signal SI;
[0155] - At least one conversion step E1, in which the conversion method as described above is applied by the conversion device 10 to the initial signal SI in order to obtain the dataset SC corresponding to the initial signal SI;
[0156] -Identification step E2, in which identification unit 5 associates the input signal SE with the signal category Ci based on the dataset SC obtained during at least one conversion step E1.
[0157] The arrangement described above enables a method for classifying an input signal SE into a signal category using a compact conversion device 10. Therefore, this classification method may include one or more conversion steps E1, depending particularly on the number of conversion devices 10.
[0158] According to a variant not shown, in this variant, classification system 1 includes N conversion devices 10 (N is an integer greater than or equal to 2), and preprocessing step E0 may include transforming the input signal SE into N initial signals SI. For example, these initial signals SI may be different or the same. Each of these initial signals SI may be associated with one of the N conversion devices 10.
[0159] Therefore, the preprocessing step E0 can be adjusted to obtain an initial signal based on the associated conversion device 10, or based on specific features to be extracted, or both. For example, the preprocessing step can be implemented based on the magnetic transmission value Rtr of each of the coupling members 30 of the N conversion devices 10.
[0160] According to this variant, the classification method may include N transformation steps E1 implemented by N transformation devices 10, and the transformation method will enable the acquisition of N datasets SC. The identification step E2 may then consist of associating the input signal SE with a signal category based on the N datasets SC obtained during the N transformation steps E1.
Claims
1. A conversion device (10) for converting an initial signal (SI) into a dataset (SC) for artificial neuron applications, the conversion device (10) comprising: - A magnetic component (20) includes a magnetic material configured to locally vary between a first magnetic state (M1) having a first magnetization and a second magnetic state (M2) having a second magnetization different from the first magnetization. The magnetic component (20) is subdivided into a plurality of regions, wherein each region is in either the first magnetic state (M1) or the second magnetic state (M2). Each region in the first magnetic state (M1) is separated from the region in the second magnetic state (M2) by a domain wall (P). - A bonding member (30) that divides the magnetic component (20) into an upstream chamber (21) and a downstream chamber (23) and provides magnetic communication between the upstream chamber (21) and the downstream chamber (23), the bonding member (30) being characterized by a magnetic transfer value (Rtr) that corresponds to the ability of the bonding member (30) to allow displacement of the domain wall (P) in a propagation direction (X), which is defined as from the upstream chamber (21) toward the downstream chamber (23). - An initialization unit (40) is configured to place the magnetic component (20) in an initial magnetic configuration (C1) in which the magnetic domain walls (P) are arranged at the bonding member (30); - Excitation unit (50), the excitation unit (50) is configured to generate physical excitation to excite the magnetic material, thereby moving the domain wall (P) in the propagation direction (X), the physical excitation being generated according to the initial signal (SI); - Detection unit (60), the detection unit (60) is arranged at the downstream chamber (23), the detection unit (60) is configured to detect and record at least one feature trigger parameter (Nbf) when a physical parameter associated with the deformation state of the domain wall (P) exceeds a predetermined threshold (Vs), the dataset (SC) including the at least one feature trigger parameter (Nbf).
2. The conversion device (10) according to claim 1 further includes a closed contour outer periphery boundary (11), in which the magnetic component (20) is completely contained.
3. The conversion device (10) according to any one of claims 1 or 2, wherein, The initialization unit (40) is configured to place the magnetic component (20) in the initial magnetic configuration (C1) when the following condition is met: at least one feature trigger parameter (Nbf) is detected by the detection unit (60).
4. The conversion device (10) according to any one of claims 1 to 3, wherein The initialization unit (40) includes an extension member (43) configured to move the magnetic domain wall (P) to the engagement member (30) so as to place the magnetic component (20) in the initial magnetic configuration (C1).
5. The conversion device (10) according to any one of claims 1 to 4, further comprising: A reconstruction unit (70) is configured to modify the magnetic transfer value (Rtr) of the joint member (30).
6. The conversion device (10) according to any one of claims 1 to 5, wherein The excitation unit (50) includes a current generator (51) configured to generate current.
7. A classification system (1) for classifying an input signal (SE) into a signal category (Ci), said classification system (1) comprising: - At least one preprocessing device (3) is designed to transform the input signal (SE) into at least one initial signal (SI); - At least one conversion device (10) according to any one of claims 1 to 6, the conversion device (10) receiving an initial signal (SI) among the at least one initial signal (SI) as input, and converting the received initial signal (SI) into a dataset (SC). - Identification unit (5), which receives the dataset (SC) as input and is configured to associate the input signal (SE) with a signal category (Ci) based on the dataset (SC), the signal category (Ci) being selected from a set of predetermined signal categories stored in the memory (7) of the classification system (1).
8. A method for converting an initial signal (SI) into a dataset (SC) for artificial neuron applications, the conversion method being implemented by a conversion device (10) according to any one of claims 1 to 6, and comprising the following stages: - Initialization phase (P1), including initial step (E13), in which the magnetic component (20) is placed in initial magnetic configuration (C1); - Excitation phase (P2), in which the excitation unit (50) generates a physical excitation based on the initial signal (SI) in order to move the magnetic domain wall (P); - Reconstruction phase (P3), which is implemented when the detection unit (60) detects that at least one physical parameter associated with the deformation state of the domain wall (P) exceeds a predetermined threshold (Vs), the reconstruction phase (P3) includes a detection step (E31), in which the detection unit (60) records at least one feature trigger parameter (Nbf) in the dataset (SC).
9. The conversion method according to claim 8, wherein, The initialization unit (40) includes a nucleating component (41), and the initialization stage (P1) includes a nucleation step (E11) performed before the initialization step (E13), in which the nucleating component (41) places the magnetic component (20) in a nucleation magnetic configuration (C0), in which all regions of the magnetic component (20) are placed in a first magnetic state (M1), except for the nucleation region placed in a second magnetic state (M2), which is contained in the upstream chamber (21).
10. The conversion method according to any one of claims 8 or 9, wherein, The reconfiguration phase (P3) includes a stop step (E32), in which the excitation unit (50) at least temporarily stops the implementation of the excitation phase (P2); a new excitation phase (P2) is implemented at the end of the reconfiguration phase (P3).
11. The conversion method according to any one of claims 8 to 10, wherein, The new initialization phase (P1) is implemented at the end of the reconstruction phase (P3).
12. The conversion method according to any one of claims 8 to 11, wherein the conversion method is performed by the conversion device (10) according to claim 5, wherein, The reconstruction stage (P3) also includes a modification step (E33) performed after the detection step (E31), in which the reconstruction unit (70) modifies the magnetic transmission value (Rtr) of the joint member (30).
13. A classification method for classifying an input signal (SE) into a signal category (Ci), said classification method being implemented by the classification system (1) according to claim 7, and comprising the following steps: - Preprocessing step (E0), in which the preprocessing device (3) transforms the input signal (SE) into at least one initial signal (SI); - At least one conversion step (E1), in which the conversion method according to any one of claims 8 to 12 is applied by the conversion device (10) to the initial signal (SI) among the at least one initial signal (SI) in order to obtain a dataset (SC) corresponding to the initial signal (SI) to which the conversion method is applied; -Identification step (E2), in which the identification unit (5) associates the input signal (SE) with the signal category (Ci) based on the dataset (SC) obtained during the at least one conversion step (E1).