• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    It is proposed an artificial neuron (30) consisting of a single-component electrical dipole comprising a single material (31) belonging to the family of Mott insulators and connected to two electric electrodes (32, 33).
    公开号:FR3020487A1
    申请号:FR1453834
    申请日:2014-04-28
    公开日:2015-10-30
    发明作者:Laurent Cario；Benoit Corraze；Pablo Stoliar；Julien Tranchant；Etienne Janod；Marie-Paule Besland；Marcelo Rozenberg
    申请人:Centre National de la Recherche Scientifique CNRS；Universite de Nantes；Universite Paris Sud Paris 11；
    IPC主号:
    专利说明:

    [0001] The field of the invention is that of neuromorphic electronic circuits, and more particularly that of artificial neurons. Artificial neural networks are inspired by biological neural networks that they mimic. More specifically, the invention relates to a new design of an artificial neuron.
    [0002] The invention has applications in particular, but not exclusively, in the field of artificial intelligence, and more particularly in the field of associative memories, image and sound recognition methods, or in the field of methods. learning or self-learning implemented by networks of artificial neurons. For example, these learning methods allow the extraction of knowledge organized from real or simulated data according to the following principle: the data considered as similar are grouped in the same group whereas the data considered as different are distributed in groups distinct. 2. TECHNOLOGICAL BACKGROUND For half a century, the information technology revolution has been closely associated with the development of the computer. However, all computers are built according to an architecture called Von Neumann. This architecture, in which the processing unit (CPU) and the memories are separate blocks, is extremely efficient for certain tasks, such as the ultra-fast processing of large data streams. However, this architecture is inefficient in many contexts and computers remain outmoded in several major classes of problems by the human brain, which operates on a very different architecture, based on a network interwoven with neurons and biological synapses.
    [0003] To overcome these limitations inherent to the Von Neumann architecture, a promising avenue is to develop a neuromorphic-type electronic architecture, made up of artificial neurons and synapses mimicking the architecture of the human brain. Essentially, the human brain is made up of a network of 1011 neurons linked together by 1015 synapses. The data are stored in the memory of the human brain through different levels of synaptic conductance, commonly called synaptic weights. Therefore, the realization of an artificial brain involves the creation of artificial neural networks that mimic the interconnection of neurons via synapses. Artificial neural networks consist essentially of neurons interconnected by synapses. For example, FIG. 1 shows a neuromorphic architecture, called a multilayer perceptron, composed of six neurons (referenced 10) and eighteen synapses (referenced 20). The synaptic weight is represented by the letter w. While Von Neumann's architecture-based microelectronics is reaching physical limits with respect to the miniaturization of electronic components, these neuromorphic architectures are generating very strong hope for researchers and industrialists, particularly with respect to energy efficiency. Various types of artificial neural networks are known in the state of the art. In most cases, these neural networks are "virtual", that is to say implemented by computer programs (called "Software Neural Networks" in English) using computers with conventional architecture, which decreases their effectiveness. , especially in terms of energy consumption. Recently, the "material" neuromorphic architectures, that is to say based on hardware neural networks (called "Hardware Neural Networks" in English), have emerged as promising alternatives, in which each neuron and each synapse is consisting of conventional electronic components. The current realization of these artificial neurons and artificial synapses requires the use of relatively complex electronic circuits and energy-consuming.
    [0004] For example, the classical implementation of an artificial synapse requires the use of a large number of components, including more than a dozen transistors.
    [0005] Recently the discovery of memristors (dipoles whose electrical resistance depends on the history of the electrical signal that passed through them) has brought an important conceptual leap in the field of artificial synapses. In comparison, the field of artificial neurons has made much less progress. An exemplary embodiment of a multi-component artificial neuron 10 according to the state of the art is illustrated in FIG. 2. This neuron 10 requires the implantation, on a printed circuit, of a large number of electronic components such as transistors (for example in CMOS technology), amplifiers and capacitors, components arranged in a relatively complex architecture.
    [0006] The implementation of different components is a relatively complex operation, many components to be implemented on a more and more restricted surface, due to the requirements of miniaturization. In addition, despite the reduction in the size of the electronic components, all still occupies a certain surface, which is difficult to reduce. In addition, such complexity greatly hampers the overall performance of the electronic circuit, in terms of integration and energy efficiency. This complexity is explained by the fact that an artificial neuron must perform several functions. First, a spiking neuron ("spiking neuron") must react to the application of electrical pulse trains. Indeed, it is in this form of impulses ("spikes") that information circulates in the most promising artificial neural networks. In addition, an artificial neuron should ideally reproduce a behavior similar to that described by the standard model LIF (for "Leaky, integrate and Fire"), that is to say a behavior implementing the three following functions: integration, flight and shooting. The principle of the LIF model is based on a simplified description of the behavior of a biological neuron, centered on the electrical charge accumulated through the pre-synaptic membrane. He models this membrane by a capacity C and a leakage resistance R in parallel. When an artificial neuron receives an input electrical signal S (t), the LIF model provides: (i) that the artificial neuron integrates ("integrate") temporally a function of this input signal; (ii) the amount corresponding to the result of this integration by the artificial neuron relaxes over time ("Leaky"). The electric signal 1r0, which results from the two phenomena integrates and leaks, is the equivalent of the potential of pre-synaptic membrane for a biological neuron; (iii) that the artificial neuron triggers a signal "Fire" when the electric signal VO reaches a given threshold. This model can be mathematically summarized by the following equation: ## EQU1 ## where: ve) represents the electrical signal after integration by the artificial neuron; RC represents the characteristic time of the exponential relaxation of the integrated signal; w represents the synaptic weight; 5 (t) is a function representing information (i.e., electrical impulses) arriving at the entrance of the artificial neuron weighted by synaptic weight w. The LIF model postulates that there is a single threshold for 4-0, independent of the shape of the input signal, beyond which the artificial LIF neuron triggers the "fire" function. In the case where the input electrical signal 5 (t) is a sequence of square electrical pulses of the same amplitude, the mathematical development of the LIF model makes it possible to establish the following theoretical relationship: NFire - X [i_ ex () 1 x [1. FIT.11] Iequation (2) - where Nare represents the number of electrical pulses needed to trigger the "fire" function; the duration of application of each electrical pulse; toif the time elapsing between two successive electrical pulses (see figure t 9):, -Fire the time required to trigger the "firing" function in the case where the electrical input signal consists of a single constant pulse of a time greater than t-Fire (see Figure 10) and T = RC the characteristic time of relaxation (see equation (1)).
    [0007] In the state of the art, patent application US 2014/0035614 (Matthew D. Pickett) is known which proposes an implementation of a "Hodgkin-Huxley" type artificial neuron by means of an electronic circuit designated under the term "neuristor". But such a circuit uses a number of electronic components such as resistors, capacitors and memristors, and is therefore built in the same spirit as the conventional artificial neurons based on CMOS transistors, discussed above. This type of neuron therefore remains complex to implement and the performance in terms of integration is limited. Moreover, this type of neuron is focused on the generation of an action potential and does not implement the three main functions of the LIF model, namely integration, flight and fire. In this context, it would be particularly interesting to be able to simplify the realization of artificial neurons and artificial neural networks. OBJECTIVES OF THE INVENTION The invention, in at least one embodiment, has the particular objective of overcoming these various disadvantages of the state of the art. More precisely, in at least one embodiment of the invention, one objective is to propose an artificial neuron of new design, which is simple to produce. At least one embodiment of the invention also aims to propose an artificial neuron that offers a strong potential for integration into electronic circuits. 4. DISCLOSURE OF THE INVENTION In a particular embodiment of the invention, there is provided an artificial neuron consisting of a single-component electrical dipole comprising a single material belonging to the family of Mott insulators connected to two electric electrodes. . Thus, the invention proposes an artificial neuron of novel design consisting of a single single-component electric dipole. The term electric dipole, an electronic component having two terminals.
    [0008] In the context of experiments carried out on Mott insulators exhibiting the volatile resistive transition phenomenon induced by electrical impulse, the inventors have discovered a new property of this family of materials that makes it possible to implement an artificial neuron in a well-controlled manner. simpler than in the prior art. Indeed, unlike the artificial neurons of the prior art which require an implementation of complex electronic circuits, the artificial single-component neuron according to the invention consists of a single electric dipole, itself made of an insulating material of Mott included between two electrodes. When this dipole is subjected to a succession of electrical pulses (representative of information from the synapses), it performs the three elementary functions of an artificial neuron as described in the reference model LIF, namely the integration with escape and shooting. This solution making it possible to implement a single elementary component that fulfills all the functionalities of an artificial neuron is all the more surprising since, for the person skilled in the art, the solutions of the prior art propose complex circuits based on various electronic components. The solutions proposed by the prior art would thus divert the person skilled in the art from the approach to be followed to achieve the invention, which goes against what has been previously established.
    [0009] Furthermore, because it implements only one electric dipole, the artificial single-component neuron according to the invention requires little energy and has a high potential for integration into the electronic circuits. According to a particular aspect of the invention, the artificial neuron comprises an input and an output. The artificial neuron is such that a first end of the electric dipole comprises a first electrical electrode which constitutes said input and a second end of the electric dipole comprises a second electric electrode which constitutes said output of the artificial neuron. According to a particular aspect of the invention, the first and second electrical electrodes are made of an electrically conductive material belonging to the group comprising: at least one of the following elements: Platinum (Pt), Gold (Au), Molybdenum (Mo ), Graphite (C), aluminum (AI), copper (Cu), doped silicon (Si); at least one of the following alloys: Brass (Cu-Zn), Steel (Fe-C), Bronze (Cu-Sn); at least one of the following transition metal compounds: TiN, TaN, RuO 2, SrRuO 3, CuO 2) - Note that this list is not exhaustive. According to a particular aspect of the invention, said material belongs to the group comprising: compounds of formula AM4Q8, with A which comprises at least one of the following elements: Ga, Ge, Zn; M which comprises at least one of V, Nb, Ta, Mo and Q which comprises at least one of S, Se, Te; compounds of formula (V1-xMx) 203, with 0 x 1, M comprising at least one of: Ti, Cr, Fe, Al, or Ga; compounds of formula NiS2_xSex, with 0 x 1. the compound of formula V02. organic Mott insulating compounds. Note that this list is not exhaustive. According to a particularly advantageous characteristic of the invention, said material is in the form of: a crystal block based on a Mott insulator; or at least one thin layer based on a Mott insulator; or a nanotube based on a Mott insulator; or a nanowire based on a Mott insulation.
    [0010] Thus, the structure of the artificial neuron can be broken down according to several particular embodiments. Thus, in the first case (material in the form of a crystal block), the structure of the artificial neuron can be likened to a three-dimensional structure. Thus, in the second case (material in the form of a thin layer), the structure of the artificial neuron can be likened to a two-dimensional structure.
    [0011] Thus, in the third and fourth cases (material in the form of a nanotube or a nano-wire), the structure of the artificial neuron can be likened to a one-dimensional structure. In another embodiment of the invention, there is provided a neural network comprising a plurality of artificial neurons interconnected by artificial synapses, said network being such that at least one artificial neuron is according to the aforementioned device (in any of its different embodiments). The invention makes it possible to drastically simplify the creation of artificial neural networks. In a particular embodiment, it is possible to envisage a network of artificial neurons, some of which are neurons according to the invention, and others are neurons of the state of the art. In another embodiment, preferential this time, each artificial neuron network is a single-component artificial neuron according to the invention.
    [0012] In another embodiment of the invention, there is provided a neuromorphic electronic circuit comprising a plurality of artificial neurons interconnected by artificial synapses or a set of electronic components, the circuit being such that at least one artificial neuron is according to the aforementioned device (in any one of its various embodiments).
    [0013] In another embodiment of the invention, there is provided a method for manufacturing an artificial neuron comprising the following steps: obtaining a material belonging to the family of Mott insulators, depositing a layer of conductive material at a first end of said Mott insulation material to form a first electrical electrode, at a second end of said Mott insulation material to form a second electrical electrode. The artificial neuron mono-component obtained according to the invention is therefore of great simplicity of manufacture. It should be noted that the deposition step of the electric electrodes can be performed either before or after the Mott insulation has been deposited.
    [0014] According to a particular embodiment, said step of obtaining a material comprises a step of cutting a crystal block belonging to the family of Mott insulators and said step of depositing a layer of conductive material is performed according to said cut crystal block.
    [0015] According to an alternative embodiment, said step of obtaining a material comprises a step of depositing, on a substrate plate, a thin layer based on a material belonging to the family of Mott insulators, said step of depositing a layer of conductive material at the first and second ends being performed according to said deposited thin layer.
    [0016] According to another variant embodiment, said step of obtaining a material comprises a step of depositing, on a substrate plate, a nanotube or nanowire based on a material belonging to the family of insulators of Mott, said step of depositing a layer of conductive material at the first and second ends being performed according to said deposited nanotube or nanowire. 5. LIST OF FIGURES Other features and advantages of the invention will appear on reading the following description, given by way of indicative and nonlimiting example, and the appended drawings, in which: FIG. 1, already described; in relation to the prior art, presents an example of a neuromorphic architecture composed of artificial neurons (illustrated by circles) connected to each other by artificial synapses (illustrated by rectangles); FIG. 2, already described in relation with the prior art, presents an electronic diagram of an artificial neuron of the state of the art; FIG. 3 shows the structure of a single-component artificial neuron according to a first particular embodiment of the invention (3D neuron); FIGS. 4A and 4B each show an exemplary structure of a single-component artificial neuron according to a second particular embodiment of the invention (2D neuron); FIGS. 5A, 5B, 5C each show an exemplary structure of a single-component artificial neuron according to a third particular embodiment of the invention (1D neuron); FIG. 6 shows the structure of a network of three mono-component artificial neurons according to a particular embodiment of the invention; FIGS. 7A, 7B, 7C, 7D present, in the form of chronograms, the evolution of electrical signals illustrating the operating principle of an artificial neuron of the integral type with leakage and pull (LIF); FIG. 8 represents an experimental circuit diagram used for the application of electrical pulses and the demonstration of the integrated behavior with leakage and pull (LIF) of the artificial mono-component neuron according to the invention; FIG. 9 represents a block diagram of a series of electrical pulses that can be applied to the experimental device described in FIG. 8; FIG. 10 presents a set of experimental curves illustrating the volatile resistive transition phenomenon induced by electrical impulse in the experimental device described in FIG. 8; FIG. 11 shows two experimental curves making it possible, on the one hand, to demonstrate the volatile nature of the resistive transition and, on the other hand, to determine the relaxation time associated with the phenomenon of leakage of the one-component artificial neuron according to FIG. 'invention. FIGS. 12A, 12B, 12C show a set of experimental curves obtained for the experimental device described in FIG. 8 and which illustrate the integrity functions with leakage and pull (LIF) of the artificial single-component artificial neuron according to the invention; FIGS. 13A, 13B show the correspondence between the experimental data obtained for the experimental device described in FIG. 8 and the theory of the LIF model. 6. DETAILED DESCRIPTION In all the figures of this document, identical elements are designated by the same numerical reference. The invention proposes a new-design monocomponent artificial neuron conforming to the reference model LIF ("Leaky integrate and Fire").
    [0017] FIG. 3 shows an exemplary structure of a single-component artificial neuron 30 according to a first particular embodiment of the invention (three-dimensional structure). The artificial single-component neuron 30 shown in this figure consists of a single electric dipole comprising a piece of material crystal 31 belonging to the family of Mott insulators, for example the compound of formula GaTa4Se8, connected by two electrodes 32 and 33. The two electrodes 32, 33 each consist of an electrically conductive material. As described in more detail below in connection with FIGS. 8 to 13, the artificial mono-component neuron thus obtained according to the invention is of integrity type with leakage and pulling, that is to say that it fills the three functions of the artificial neuron conform to the reference model LIF. In this example, the insulating material of Mott 31 is a cut crystal block of thickness 20 μm, length 300 μm and width 200 μm. Each electrical electrode is typically in the form of a thin layer of 0.1 μm thickness, length 300 μm and width 200 μm. It should be noted that these dimensions are given for illustrative purposes only and may of course be different. In this particular embodiment, the crystal block 31 being of three-dimensional (3D) structure, the neuron 30 according to the invention is likened to a three-dimensional (3D) artificial neuron.
    [0018] In general, the dimensions of the Mott 31 insulating crystal piece, the electrodes 32, 33 and their arrangement with respect to said piece of crystal, as well as the choice of materials can be optimized so that the mono artificial neuron -composing 30 thus obtained can benefit from the best performance (quality of the electric dipole response compared to that expected in the LIF model of the artificial neuron, structural integration, etc.).
    [0019] The main steps of the method of manufacturing the single-component artificial neuron 30 shown in FIG. 3 are described below. Firstly, a piece of GaTa4Se8 crystal constituting the functional material of the mono artificial neuron is cut out. -composing 30. Then is deposited on the two opposite sides of the piece of crystal 31, a conductive material constituting the input electrodes 32 (denoted "INPUT") in the figure) and output 33 (denoted " OUTPUT "in the figure) of the artificial neuron. This deposit can be made typically either by applying directly to the faces of the crystal a conductive lacquer or using a deposition technique such as Joule evaporation or magnetron sputtering (PVD for "Physical Vapor Deposition"). FIGS. 4A and 4B show two examples of the structure of a single-component artificial neuron 30 according to a second particular embodiment of the invention (two-dimensional structure).
    [0020] The artificial single-component neuron 40A represented in FIG. 4A consists of a single electric dipole comprising a layer of material 41A belonging to the family of Mott insulators, for example the compound of formula GaV4S8, connected by two electrically conductive electrodes 42A. and 43A. The electric dipole is supported by an insulating substrate 44A. It is therefore the electric dipole 40A as simple elementary electronic component that behaves like an artificial neuron, the insulating substrate 44A only playing the role of support of the device of the invention. As described in more detail below with reference to FIGS. 8 to 13, the artificial single-component artificial neuron 40A thus obtained according to the invention is of integrity type with leakage and pulling, that is to say that it fills the three functions of the artificial neuron conform to the reference model LIF. In the embodiment illustrated here, the first electrode 42A is in the form of an L is disposed partly on a first end of the Mott insulation layer 41A, the other part extending on the substrate insulation 44A. This electrode forms the input of the artificial single-component neuron 40A (denoted "INPUT") in the figure). The second electrode 43A is in the form of an L and is partly disposed on the second end of the Mott insulation layer 41A, the other portion extending on the insulating substrate 44A. This electrode forms the output of the single-component artificial neuron 40A (denoted "OUTPUT" in the figure). It is noted here that the set of electrodes 42A, 43A is disposed on the Mott insulation layer 41A, in other words once the Mott insulation layer 41A has been deposited on the substrate 44A. This is of course an example of a particular structure. It is clear that many other single-component artificial neuron structures can be envisaged, without departing from the scope of the invention. In particular, it is possible to provide for example a neuron structure according to which the electrodes are arranged under the Mott insulator layer 41A (as is the case in the particular embodiment illustrated in FIG. 6). It is also possible, for example, for the Mott insulation layer to be sandwiched between the electrodes, as shown in FIG. 4B. Likewise, the shape and dimensions of the electrodes may also vary depending on the desired neuromorphic architecture. The artificial single-component neuron 40B shown in FIG. 4B consists of a single electric dipole comprising a layer of material 41B, of formula GaV4S8, sandwiched between the electrodes 42A and 43A. The whole is arranged on a substrate plate 44B. In these two examples, the layer of insulating material of Mott is a thin layer of thickness 0.1 μm, length 20 μm and width 1 μm. Each electrode has a thickness of 0.1 μm. The term "thin layer" in the remainder of this document, a layer of material whose thickness is generally less than 10 microns, as opposed to "thick layers" whose thickness is generally greater than 10 microns. It should be noted that these dimensions are given for illustrative purposes only and may of course be different. In this particular embodiment, it is considered that the thin layer of Mott insulating material has a two-dimensional (2D) structure. The 40A or 40B neuron according to the invention can therefore be likened to a two-dimensional (2D) artificial neuron.
    [0021] In general, the dimensions of the Mott insulation layer, the metal electrodes and their arrangement with respect to said layer, as well as the choice of materials can be optimized so that the single-component artificial neuron thus obtained can benefit from the best performances (quality of the response of the electric dipole compared to that expected in the LIF model of the artificial neuron, structural integration, etc.).
    [0022] The following are the main steps of the manufacturing process of the single-component artificial neuron 40A shown in FIG. 4A. Firstly, a layer of GaV4S8 constituting the functional material of the artificial single-component neuron 40A is deposited on a substrate plate (an oxidized silicon wafer for example). This deposition can be carried out typically by means of a deposition technique such as, for example: Joule evaporation, magnetron sputtering (PVD), laser pulse deposition (PLD), deposition by atomic layer (ALD for "Atomic Layer Deposition"), the deposit by chemical solutions (CSD for "Chemical solution deposition"), the deposition by screen printing, deposit by spin / dip ("Spin / Dip coating" in English). Then, we proceed to the deposition of a metallic material constituting the input electrodes 42A and output 43A of the neuron, as explained above in connection with FIG. 3. For the artificial single-component neuron 40B, we proceed from first depositing, on a substrate plate (an oxidized silicon wafer for example), a layer of metal material constituting the output electrode 43B according to a technique of Joule evaporation or magnetron sputtering, for example. Then, a layer of GaV4S8 constituting the functional material 41B of the artificial single-component artificial neuron 40B is deposited on said layer of metallic material 43B by means of one of the deposition techniques described above in connection with FIG. 4A. . Finally, another layer of metallic material constituting the input electrode 42B of the neuron is deposited. FIGS. 5A and 5B show two examples of structure of a monocomponent artificial neuron according to a third particular embodiment of the invention (one-dimensional structure).
    [0023] The single-component artificial neuron 50A shown in FIG. 5A consists of a single electric dipole comprising a Mott 51A insulating nanowire of formula (V1, Crx) 203, connected by two electrodes 52A and 53A. The electric dipole is supported by an insulating substrate 54A. It is therefore the electric dipole 50A as simple elementary electronic component that behaves like an artificial neuron, the insulating substrate 54A only playing the role of support of the device of the invention. As described in more detail below with reference to FIGS. 8 to 13, the artificial single-component artificial neuron 50A thus obtained according to the invention is of integrity type with leakage and pulling, that is to say that it fills the three functions of the artificial neuron conform to the reference model LIF. In the embodiment illustrated here, the first electrode 52A is disposed in part under a first end of the Mott 51A insulator nanowire, the other portion extending on the insulating substrate 54A. This electrode forms the entrance of the single-component artificial neuron 50A (denoted "INPUT") in the figure). The second electrode 53A is disposed partially under the second end of the Mott 51A nano-insulator wire, the other portion extending on the insulating substrate 54A. This electrode forms the output of the single-component artificial neuron 50A (denoted "OUTPUT" in the figure). It should be noted here that the set of electrodes 52A, 53A is disposed under the Mott 51A insulating nano-wire, that is, before the Mott 51A insulating nano-wire has been deposited on the substrate 54A. This is of course an example of a particular structure. It is clear that many other single-component artificial neuron structures can be envisaged, without departing from the scope of the invention. The following are the main steps of the manufacturing process of the single-component artificial neuron 50A shown in FIG. 5A.
    [0024] Firstly, a layer of metallic material constituting the two input and output electrodes of the neuron is deposited on a substrate plate (a silicon wafer for example). This deposition may be carried out typically using one of the following techniques, for example: Joule evaporation, magnetron sputtering (PVD), laser pulse deposition (PLD), atomic layer deposition (ALD), chemical solution deposition ( CSD), silk screening, spin coating / dip coating ("Spin / Dip coating"). Then, the nano-wire is positioned so that it touches both the electrodes 52A and 53A, the nano-wire having been previously synthesized by means of a vapor-liquid-solid method (VLS). or template for example. The example of FIG. 5B differs from that of FIG. 5A in that the artificial neuron 50B consists, not of a nano-wire of Mott insulation, but of a nano-tube of Mott insulation. 51B. The rest of the structure of this artificial neuron 50B is identical to the artificial neuron 50A. An input conductive electrode 52B is arranged under the nano-tube 51B at one of its ends and an output conductive electrode 53B is arranged under the nano-tube 51B at the other of its ends, the all being disposed on an insulating substrate 54B. The artificial neuron 50C shown in FIG. 5C consists of an electric dipole comprising a Mott 51C insulating nanowire of formula V2_xCrx03, connected by two electrodes 52C and 53C. In this third example, the electric dipole is supported by a generally L-shaped insulating substrate 54C having two substantially parallel surfaces 56 and 57 on which the layers of material forming the 52C and 53C output electrodes respectively are arranged respectively. artificial neuron. The two surfaces 56 and 57 are connected by a third surface 58 on which the nanowire 51C is arranged so that the end sections thereof are connected to the 52C input and 53C output electrodes of the neuron. First of all, a layer of metallic material constituting the output electrode 53C of the neuron is deposited on the L-shaped substrate at the level of the surface 57, for example by Joule evaporation. Then, the nanowire 51C is arranged so that one of its ends is in contact with the output electrode 53C and so that a part of the circumferential surface of the nanowire extends along the surface 58 in the direction of the surface 56, the nano-wire having been previously synthesized by means of a vapor-liquid-solid method (VLS) or template for example. Finally, another layer of metallic material is deposited at the free end of the nano-wire and the surface 56 to constitute the input electrode 52C of the neuron, for example by Joule evaporation.
    [0025] In this particular embodiment, it is considered that the nano-wire or nanotube insulation Mott has a one-dimensional structure (1D). The 50A, 50B or 50C neuron according to the invention can therefore be likened to a one-dimensional (1D) artificial neuron.
    [0026] In the artificial neuron as described above, in any of its various embodiments (1D, 2D, 3D), the input and output electrodes can be made in one of the following materials: a single element such as Platinum (Pt), Gold (Au), silver (Ag), Molybdenum (Mo), Graphite (C), Aluminum (AI), Copper (Cu), doped silicon (Si)) or an alloy such as brass, steel, bronze for example or a transition metal compound such as TiN, TaN, RuO 2, SrRuO 3, Cu52. For integration into an artificial neural network, each electrode of the single-component artificial neuron as previously described, in any one of its various embodiments (1D, 2D, 3D), is configured to be connected at one or more artificial synapses (not shown in the figure), such as artificial synapses w according to the neuromorphic architecture shown in FIG. 1 (where each artificial neuron is configured to be connected to a set of three input synapses and three synapses in output). The neural network according to the invention comprises a plurality of artificial neurons interconnected by artificial synapses. In a particular embodiment ("degraded" mode), it is possible to envisage a network of artificial neurons, some of which are artificial mono-component neurons according to the invention, and others are neurons of the state of the art. In another embodiment (preferential mode), each artificial neuron of the network is a single-component artificial neuron as defined according to the invention. FIG. 6 shows a network of three mono-component artificial neurons 601, 602, 603, according to a particular embodiment of the invention. The artificial single-component neuron 601 is formed of an electric dipole comprising a thin layer 611 of GaV4S8 material whose ends are brought into contact with a set of metallic electrodes acting as input / output of the neuron: the first end of said GaV4S8 layer is contacted with the input electrode 621 and the second end of said GaV4S8 layer is contacted with the output electrode 631. The artificial single-component neuron 602 is formed of an electric dipole comprising a thin layer 612 of GaV4S8 material whose ends are brought into contact with a set of metal electrodes acting as input / output of the neuron: the first end of said GaV4S8 layer is contacted with the input electrode 622 and the second end of said GaV4S8 layer is contacted with the output electrode 632. The one-component artificial neuron 603 is formed of an electric dipole comprising a thin layer 613 of GaV4S8 material whose ends are brought into contact with a set of metal electrodes acting as input / output of the neuron: the first end of said GaV4S8 layer is contacted with the input electrode 623 and the second end of said GaV4S8 layer is contacted with the output electrode 633.
    [0027] Each single-component artificial neuron was obtained by means of the manufacturing method described above in connection with FIG. 3 on an insulating substrate plate 64, with the difference that the input / output electrodes here are of straight elongated shape and that they are placed under the GaV4S8 thin layers. It is conceivable that these single-component artificial neurons could be connected to artificial synapses (not shown in the figure) at the input (INPUT) and output (OUTPUT) of each neuron so as to integrate a neural network. artificial (as the example of neuromorphic architecture shown in Figure 1 where each neuron is connected to a set of three input synapses and three synapses output).
    [0028] Artificial neural networks can also be envisaged with, on the one hand, a number of artificial neurons greater than three, and on the other hand comprising artificial neurons either based on pieces of Mott insulator crystal, or on the basis of Thin layer of Mott insulation, either based on nanowire or Mott insulation nanotube.
    [0029] FIGS. 7A, 7B, 7C present, in the form of chronograms, the evolution of electrical signals illustrating the operating principle of an artificial neuron of the integral type with leakage and pull (LIF). An artificial LIF neuron receives a synaptic weight-weighted pulse sequence S (t) w (Figure 7A). The artificial neuron integrates ("integrate") then temporally a function of this input signal; the amount corresponding to the result of this integration by the artificial neuron relaxes over time ("Leaky") thus giving the signal v (t) (Figure 7B). This signal v (t) is the equivalent of the pre-synaptic membrane potential for a biological neuron. The artificial neuron triggers an output signal ("Fire") when the signal V (t) reaches a given threshold. This output signal S0 (t) (Figure 7C) is a pulse whose shape is not explicitly defined in the LIF model. The artificial neuron according to the LIF model thus fulfills the three following functions: integration with leakage and shooting. The experimental part described below shows that the Mott insulators have the properties required to behave as an artificial neuron according to the LIF model. FIG. 8 represents an experimental circuit diagram used for the application of electrical pulses and the demonstration of the integrity with leakage and pull (LIF) behavior of the artificial single-component artificial neuron according to the invention.
    [0030] It comprises an electric pulse generator 83 connected in series with an experimental device consisting of a conductance denoted "w" 84 acting as an artificial synapse itself in series with a piece of crystal of functional material 80 arranged between two metal electrodes 81 and 82. The objective was to test the response of the functional material 80 to the application of different series of electrical pulses emitted by the generator 83. The functional material 80 is here a piece of crystal GaTa4Se8 which was previously synthesized, then cleaved. It is one of the compounds belonging to the family of Mott AM4Q8 insulators, with A which comprises at least one of the following elements: Ga, Ge, Zn, M which comprises at least one of the following elements: : V, Nb, Ta, Mo and Q which comprises at least one of the following elements: S, Se, Te.
    [0031] The functional material 80 typically has the following dimensions: 300 μm long, 200 μm wide and 20 μm thick. The two metal electrodes 81 and 82 are formed for example by means of a carbon lacquer. The voltage across the conductance "W"> 84 (denoted Vw) and the voltage across the functional material 80 (denoted V) are measured using an oscilloscope. The voltage of the pulses is defined according to the relation Vtota, = V + Vw and the resistance of the functional material (denoted R) is calculated according to Ohm's law: R = VA, where / = w V. To simplify the assembly in practice the functional material 80 receives electrical signals only from the single pulse generator 83. However, there is no obstacle for this functional material to be connected to several different sources of pulses, in order to mimic the large number of of synaptic connections that may exist in an artificial neural network. FIG. 9 diagrammatically illustrates the main characteristics of the series of electrical pulses sent by the generator: the duration t0 of the electrical pulses, the duration toFF flowing between two successive pulses, the voltage Vtotai of each pulse and the number N of pulses emitted in the series. FIG. 10 shows the response of the functional material 80 when the circuit of FIG. 8 is subjected to a single electrical pulse of long duration (greater than 100 μl) and of amplitude of approximately 60 V. This response is here characterized by the voltage V obtained at the terminals of the functional material 80, the current I passing through it and its electrical resistance R. After a period of about 70 lis (t Eire = 70 lis) after the start of the electrical pulse, a sudden voltage drop V is observed at the terminals of the functional material 80, concomitant with a strong increase in the current / which passes through it (see the characteristics / (t) and V (t) shown in FIG. 10), and therefore a sharp fall resistance. These experimental curves highlight the physical phenomenon of resistive transition from an electric field threshold. For the GaTa4Se8 compound, the value of the threshold electric field is approximately equal to 2.5 kV / cm, which corresponds to a threshold voltage (V ',,',) equal to 10V in this example. The threshold value of the electric field varies in power law with the band gap energy of Mott insulating compounds and can therefore be determined according to this law. It is therefore possible by the choice of the Mott insulation used to optimize the value of this threshold field according to the desired characteristics. FIG. 11 shows two experimental curves obtained for the functional material of the experimental setup of FIG. 8, making it possible to illustrate that the resistance of the functional material returns to its initial state after the pulse and therefore that the resistive transition is volatile. These experimental curves were obtained under the same conditions as those described in FIGS. 10, 12A-C. The first graph shows the evolution of the voltage Vtota I applied to the functional material 80 and the second graph shows the evolution of its resistance over time. It can be seen that the resistance R of the functional material 80, after the resistive transition has been completed (that is to say after the sharp drop in resistance due to the electric pulse), relaxes until it regains its state. of equilibrium according to a decreasing exponential law. The refinement of the data shown in the insert A of FIG. 10 makes it possible to determine the characteristic relaxation time: ## EQU1 ## The application of electrical pulses thus makes it possible to vary, in a volatile manner, the resistance of this material between at least two distinct resistance states. On the basis of this observation, the inventors have discovered that this property of Mott insulators can be exploited to implement a single-component artificial neuron of the Integral Type with Leak and Pull (LIF). 12A, 12B, 12C, a set of experimental curves obtained for the functional material 80 which illustrate the feasibility of the integrity functions with leakage and pull (LIF) of the artificial single-component artificial neuron according to the invention are now presented in relation with FIGS. .
    [0032] FIG. 12A shows the response of the functional material 80 (the output signal is the intensity of the electric current /), when the circuit of FIG. 8 is subjected to a series of six electrical pulses of duration n = 15, separated by each other having a duration of TO = 30 μs and an amplitude of approximately 60 V (Vtota input signal After applying a series of six electrical pulses to the functional material 80, a sudden increase in the current I passing through the material The sudden appearance of this sharp increase in current I when the sixth electrical pulse is applied shows that a signal ("spike") has been triggered by the functional material 80. This result is very important since proves that the functional material realizes on the one hand the function "shooting", and on the other hand, an integration of a function of the input signal Vtote Indeed, each electrical pulse has not, alone, of notable effect since its duration is lower than tf, '= 70 they (see Figure 10), while several successive electrical pulses have the effect of inducing the phenomenon of "shooting". This result therefore shows the realization of the "fire" (or "Fire" in English) and "integration" (or "integrate" in English) functions, that is to say two of the three functions required by the LIF model. The phenomenon of integration is also confirmed by a series of experiments carried out by keeping the duration toFF between two successive successive pulses and by varying the duration toN of the electrical pulses. FIG. 12B shows the response of the functional material 80 (output signal /), when the circuit of FIG. 8 is subjected to a series of four electrical pulses with a duration of 20 N, separated from each other by a duration of 30 ls, and amplitude of about 60 V (Vtota / input signal). Thus, FIG. 12B shows the same experiment as that described in FIG. 12A, but with a duration t0 of electrical pulses applied to the augmented functional material (a change in duration from 15 to 20 μs) is obtained. It can be seen here that the number of pulses necessary for the functional material 80 to trigger an output signal ("spike") decreases as the duration of the electrical pulses increases (from 6 to 4 pulses required). This again makes it possible to show that the experimental device is capable of implementing the functions "Integration" ("Integrate") and "Shooting" ("Fire"), that is to say two of the three functions of the standard model. LIF. Figure 12C shows the same experiment as that shown in Figure 12A, but with a duration toFF between two successive pulses higher (170 against 30 they previously). The result is very clear: the number of electrical pulses necessary for the functional material 80 to trigger an output signal ("spike") has increased from six to eight pulses. This proves that the signal integrated by the functional material relaxes over time, which reflects a leaky form. This makes it possible to show that the functional material is capable of implementing, not only the integration ("Integrate") and fire ("Fire") functions, but also the leakage function ("Leaky") of the standard model. LIF.
    [0033] FIG. 13 shows the experimental dependence of the number of pulses necessary to draw (NF, ') with the duration t' (FIG. 13A) and the duration te (FIG. 13B) obtained during experiments similar to those described in FIGS. 12A. -VS. There is a very good agreement between the experimental points and the theoretical relation: TI x [1 tFA 1 equation (2) ff - obtained by the mathematical development of the LIF model in the case of square electric pulses (whose principle is described more high in relation to the state of the art). This very good agreement is all the more remarkable because the theoretical dependence contains no adjustable parameter since the duration t -Fire and the relaxation time r have been fixed at the experimental values (tire 70us and r = 610 us). This demonstrates that the mathematical prediction of the LIF model (equation (2)) applies to the Mott insulating single-component artificial neuron of the invention. This very good agreement proves that there exists a single threshold of integration, independent of the durations t 'and te, beyond which the artificial neuron according to the invention triggers the firing function.
    [0034] All of these experimental results thus prove that the Mott GaTa4Se8 insulation-based experimental device 80 performs the three essential functions of the artificial neuron described by the LIF model ("Leaky", "Integrate", "Fire"), know integration with leakage and firing beyond a threshold. In addition, these results show that it is possible to predict, for the artificial neurons according to the invention, the number of electrical pulses needed to generate the "firing" function. The artificial mono-component neuron described above, in any of its various embodiments (1D, 2D, 3D), is based on the use of a Mott insulator of formula GaTa4Se8 or GaV4S8. It is clear that any other material belonging to the family of Mott insulators can be considered as functional material of a single-component artificial neuron without departing from the scope of the invention, for example: compounds of formula AM4Q8; with A which includes at least one of: Ga, Ge, Zn, M which comprises at least one of V, Nb, Ta, Mo and Q which includes at least one of the following : S, Se, Te; compounds of formula (V1-xMx) 203, with 0 x 1, M comprising at least one of: Ti, Cr, Fe, Al, or Ga; compounds of formula NiS2_xSex, with 0 x 1. the compound of formula V02. organic Mott insulating compounds.
    权利要求:
    Claims (11)
    [0001]
    REVENDICATIONS1. Artificial neuron (30) characterized in that it consists of a single-component electric dipole comprising a single material (31) belonging to the family of Mott insulators connected to two electric electrodes (32, 33).
    [0002]
    An artificial neuron according to claim 1, including an input and an output, and wherein a first end of the electric dipole (30) comprises a first electrical electrode (32) which constitutes said input and a second end of the electric dipole comprises a second electric electrode (33) which constitutes said output of the artificial neuron.
    [0003]
    The artificial neuron according to claim 2, wherein the first and second electrodes (31, 32) are made of an electrically conductive material belonging to the group comprising: at least one of the following elements: Platinum (Pt), Gold (Au), molybdenum (Mo), graphite (C), aluminum (AI), copper (Cu), doped silicon (Si); at least one of the following alloys: Brass (Cu-Zn), Steel (Fe-C), Bronze (Cu-Sn); at least one of the following transition metal compounds: TiN, TaN, RuO 2, SrRuO 3, CuO 2) -
    [0004]
    4. An artificial neuron according to any one of claims 3, wherein said material belonging to the family of Mott insulators belongs to the group comprising: compounds of formula AM4Q8, with A which comprises at least one of the following elements: Ga, Ge, Zn; M which comprises at least one of V, Nb, Ta, Mo and Q which comprises at least one of S, Se, Te; compounds of formula (V1-xMx) 203, with 0 x 1, M comprising at least one of: Ti, Cr, Fe, Al, or Ga; compounds of formula NiS2_xSex, with 0 x 1.the compound of formula V02. organic Mott insulating compounds.
    [0005]
    An artificial neuron according to any one of claims 4 to 4, wherein said material is in the form of a thin layer based on a Mott insulator; or in the form of a crystal block based on a Mott insulator; or in the form of a nanotube based on a Mott insulator; or in the form of a nanowire based on an insulating material of Mott.
    [0006]
    6. A neural network comprising a plurality of artificial neurons interconnected by artificial synapses, said network being characterized in that at least one artificial neuron is according to any one of claims 1 to 5.
    [0007]
    7. Neuromorphic electronic circuit comprising a plurality of artificial neurons interconnected by artificial synapses or a set of electronic components, said circuit being characterized in that at least one artificial neuron is according to any one of claims 1 to 5.
    [0008]
    8. A method of manufacturing an artificial neuron characterized in that it comprises the following steps: obtaining a material belonging to the family of Mott insulation, depositing a layer of conductive material: * at a level of first end of said Mott insulator material to form a first electrical electrode, at a second end of said Mott insulator material to form a second electrical electrode.
    [0009]
    The manufacturing method according to claim 8, wherein said step of obtaining a material comprises a step of cutting a crystal block belonging to the family of Mott insulators, and wherein said deposition step of a layer of conductive material is performed as a function of said cut crystal block.
    [0010]
    The manufacturing method according to claim 8, wherein said step of obtaining a material comprises a step of depositing, on a substrate plate, a thin layer based on a material belonging to the family of insulants. de Mott, and wherein said step of depositing a layer of conductive material at the first and second ends is performed according to said deposited thin layer.
    [0011]
    The manufacturing method according to claim 8, wherein said step of obtaining a material comprises a step of depositing, on a substrate plate, a nanotube or nanowire based on a material belonging to the family of Mott insulators, and wherein said step of depositing a layer of conductive material at the first and second ends is performed according to said deposited nanotube or nanowire.
    类似技术:
    公开号 | 公开日 | 专利标题
    EP3138050B1|2021-05-05|Single-component artificial neuron made from mott insulators, network of artificial neurons, and method for making said artificial neurons
    Wang et al.2018|Capacitive neural network with neuro-transistors
    Del Valle et al.2018|Challenges in materials and devices for resistive-switching-based neuromorphic computing
    Pantazi et al.2016|All-memristive neuromorphic computing with level-tuned neurons
    EP2965269B1|2018-08-22|Artificial neuron and memristor
    EP2713318B1|2019-06-12|Neuromorphic system exploiting the intrinsic characteristics of memory cells
    EP1506581B1|2009-02-11|Superconducting quantum bit device with josephson junctions
    FR2977350A1|2013-01-04|NETWORK OF ARTIFICIAL NEURONS BASED ON COMPLEMENTARY MEMRISTIVE DEVICES
    US20190303744A1|2019-10-03|Capacitive artificial neural networks
    US20150154469A1|2015-06-04|Pattern recognition method and apparatus for the same
    Lee et al.2020|Simple artificial neuron using an ovonic threshold switch featuring spike-frequency adaptation and chaotic activity
    Kuncic et al.2021|Neuromorphic nanowire networks: principles, progress and future prospects for neuro-inspired information processing
    CN109585650B|2020-05-05|Glial cell-like neuromorphic device and preparation method thereof
    CN107909146A|2018-04-13|Neuron circuit based on volatibility threshold transitions device
    FR2946788A1|2010-12-17|DEVICE WITH ADJUSTABLE RESISTANCE.
    EP3660749A1|2020-06-03|Neural circuit suitable for implementing synaptic learning
    EP2122704B1|2010-11-24|Use of lacunar spinels with tetrahedral aggregates of a transition element of the of the am4x8 type in an electronic data rewritable non volatile memory, and corresponding material
    EP0938732B1|2001-07-11|Analog electronic cell, in particular designed to be implanted in an integrated circuit, and integrated circuit comprising such cells
    WO2022053502A1|2022-03-17|Device implementing a convolution filter of a neural network
    EP3718054A1|2020-10-07|Neuromimetic network and related production method
    EP3827377A1|2021-06-02|Neural network comprising spintronic resonators
    EP3827378A1|2021-06-02|Synaptic chain comprising spintronic resonators based on the inverse spin hall effect and neural network comprising such a synaptic chain
    WO2021156419A1|2021-08-12|Classification device, classification set and associated classification method
    FR3084503A1|2020-01-31|SYNAPTIC CHAIN COMPRISING SPINTRONIC RESONATORS BASED ON THE EFFECT OF SPIN DIODE AND NEURON NETWORK COMPRISING SUCH A SYNAPTIC CHAIN
    EP3053107A1|2016-08-10|Neuromimetic circuit and method of fabrication
    同族专利:
    公开号 | 公开日
    CN106462799A|2017-02-22|
    CN106462799B|2019-12-31|
    US20170124449A1|2017-05-04|
    WO2015165809A3|2015-12-30|
    WO2015165809A2|2015-11-05|
    US10867237B2|2020-12-15|
    EP3138050A2|2017-03-08|
    EP3138050B1|2021-05-05|
    FR3020487B1|2017-10-06|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    RU2627125C1|2016-10-24|2017-08-03|федеральное государственное автономное образовательное учреждение высшего образования "Московский физико-технический институт "|Nano-sized artificial neuron "integrate-and-fire"|WO2000041245A1|1998-12-30|2000-07-13|Alexander Mikhailovich Ilyanok|Quantum-size electronic devices and methods of operating thereof|
    US7303971B2|2005-07-18|2007-12-04|Sharp Laboratories Of America, Inc.|MSM binary switch memory device|
    CN101364594B|2007-08-09|2010-06-02|中国科学院半导体研究所|Silicon based single electron neure quantum circuit|
    FR2989212B1|2012-04-10|2014-05-02|Centre Nat Rech Scient|USE OF CENTROSYMETRIC MOTOR INSULATION IN A RESISTIVE SWITCHING DATA STORAGE MEMORY|
    US8669785B2|2012-07-31|2014-03-11|Hewlett-Packard Development Company, L.P.|Logic circuits using neuristors|
    US9165246B2|2013-01-29|2015-10-20|Hewlett-Packard Development Company, L.P.|Neuristor-based reservoir computing devices|TWI625681B|2017-05-11|2018-06-01|國立交通大學|Neural network processing system|
    CN107563505A|2017-09-24|2018-01-09|胡明建|A kind of design method of external control implantation feedback artificial neuron|
    CN109754070A|2018-12-28|2019-05-14|东莞钜威动力技术有限公司|Insulation resistance value calculation method neural network based and electronic equipment|
    CN110411850B|2019-07-22|2020-10-09|北京科技大学|Method for evaluating service conditions of high-temperature alloy turbine blade|
    法律状态:
    2015-03-19| PLFP| Fee payment|Year of fee payment: 2 |
    2015-10-30| PLSC| Publication of the preliminary search report|Effective date: 20151030 |
    2016-03-23| PLFP| Fee payment|Year of fee payment: 3 |
    2017-03-22| PLFP| Fee payment|Year of fee payment: 4 |
    2018-04-26| PLFP| Fee payment|Year of fee payment: 5 |
    2019-04-29| PLFP| Fee payment|Year of fee payment: 6 |
    2020-04-30| PLFP| Fee payment|Year of fee payment: 7 |
    2021-04-29| PLFP| Fee payment|Year of fee payment: 8 |
    优先权:
    申请号 | 申请日 | 专利标题
    FR1453834A|FR3020487B1|2014-04-28|2014-04-28|MONO-COMPONENT ARTIFICIAL NEURONE BASED ON MOTT INSULATORS, ARTIFICIAL NEURON ARRAY AND METHOD FOR MANUFACTURING THE SAME|FR1453834A| FR3020487B1|2014-04-28|2014-04-28|MONO-COMPONENT ARTIFICIAL NEURONE BASED ON MOTT INSULATORS, ARTIFICIAL NEURON ARRAY AND METHOD FOR MANUFACTURING THE SAME|
    PCT/EP2015/058873| WO2015165809A2|2014-04-28|2015-04-24|Single-component artificial neuron made from mott insulators, network of artificial neurons, and method for making said artificial neurons|
    CN201580023093.2A| CN106462799B|2014-04-28|2015-04-24|Single-part artificial neuron based on Modrin insulator, artificial neuron network and corresponding manufacturing method|
    EP15718863.2A| EP3138050B1|2014-04-28|2015-04-24|Single-component artificial neuron made from mott insulators, network of artificial neurons, and method for making said artificial neurons|
    US15/307,269| US10867237B2|2014-04-28|2015-04-24|Single-component artificial neuron based on Mott insulators, network of artificial neurons and corresponding manufacturing method|
    [返回顶部]