JP6985993B2 - Learning method of electronic circuit and neural network using it - Google Patents

Description

The present invention relates to artificial intelligence or a machine learning technique thereof.

Conventionally, a technique called a quantum dot (Quantum Dot, abbreviated as QD) is known. For example, in a semiconductor, when a granular structure having a size corresponding to the de Broglie wavelength of an atom is formed by crystal growth or microfabrication, electrons are confined in that region. Generally, a quantum dot is a quantum dot in which electrons or holes are confined from all three-dimensional directions (for example, Non-Patent Document 1).

In recent years, technologies such as artificial neurons, neural networks, and machine learning for artificial intelligence have been actively studied (for example, Non-Patent Document 2).

Non-Patent Document 3 is a document relating to a current mirror element referred to in Examples described later.

M. Sugawara, “Self-Assembled InGaAs / GaAs Quantum Dots”, Semiconductors and Semmimetals, Vol. 60 (1999). P. Smolensky, M. C. Mozer, D. E. Rumelhart, “Mathematical Perspectives on Neural Network,” Lawrence Erlbaum Associates (1996). D. Nandi, A. D. K. Finck, J. P. Eisenstein, L. N. Pfeiffer and K. W. West, "Exciton condensation and perfect Coulomb drag", Vol. 488, pp.481-.484 (2012).

Artificial intelligence is a technology that realizes functions such as human intelligence on a computer by means of a neural network in which artificial neurons are connected in a network with reference to the neural structure of the brain. With artificial intelligence, for example, it is possible to determine whether the subject of a photograph is a cat or a dog, or to transcribe characters from voice. In order to exert these functions, it is necessary to adjust the connection strength between artificial neurons in the neural network in advance and set it to the optimum value. Currently, neural networks with about 1 million artificial neurons and 1 billion connections between artificial neurons have been implemented, but with the increase in the scale of neural networks, the above-mentioned connections have been implemented. The time required to search for the optimum value of intensity (called machine learning) has become a problem.

As a solution approach to this problem, there are improvements in the algorithm for finding the optimum value and parallelization of processing by hardware. The former is known to use a genetic algorithm or a steepest descent method. For the latter, parallel computing using a processor that is good at parallel processing such as a graphics processing unit (GPU) is known. Both have a certain effect on shortening the calculation time, but have not yet solved the problem of increasing the exponential calculation time with respect to the scale of the neural network.

In general, while it is easy to obtain the output from the input and the bond strength in a neural network, it is difficult to analytically obtain the bond strength from the input and the output, which is a kind of inverse problem. Therefore, in the existing machine learning, the optimum value of the bond strength must be searched by trial and error so that the desired output can be obtained. This is a factor in the exponential increase in the time required for machine learning with respect to the scale of the neural network. Although the speed has been increased to some extent by improving algorithms such as backpropagation, it is still in the category of exponential problems (for example, Non-Patent Document 2).

Since it is considered difficult to improve the efficiency of machine learning using only mathematical methods, a completely different approach is expected to find the optimum value of the coupling strength of the neural network.

An object to be solved by the present invention is to provide a technique capable of efficiently specifying the coupling strength in the learning of an artificial neural network.

A preferred aspect of the present invention is an electronic circuit including quantum dots, a capacitance section, and an electrical resistance section. In this electronic circuit, the quantum dot comprises a first electrode, a second electrode, and a third electrode, the first electrode is connected to the first potential, and the second electrode is the first current source. Connected, the third electrode is connected to the second current source. The structure is such that the amount of current of electrons or holes flowing in the electric resistance portion can be measured.

Another preferred aspect of the present invention is a method for learning a neural network using an electronic circuit including quantum dots, a capacitance section, and an electrical resistance section. In this electronic circuit, the quantum dot includes a first electrode, a second electrode, and a third electrode, and the first electrode is connected to the first potential, and the amount of current of electrons or holes flowing in the electric resistance section. It has a structure that can measure. In the learning of the neural network, the first step of setting the current value corresponding to the problem of the teacher data to the first current source connected to the second electrode, and the second current source connected to the third electrode. The second step of setting the current value corresponding to the answer of the teacher data, the third step of measuring the current amount of the electron or hole flowing in the electric resistance part, and the fourth step of constructing the neural network corresponding to the electronic circuit. , A fifth step of setting a value corresponding to the amount of current as the coupling strength of the neural network is included.

On both sides, in a more specific and preferred embodiment, electrons or holes are stably flowed from the first potential through the quantum dots to the first electrode and the second electrode, and the quantum dots and the second electrode are connected. The amount of electron or hole current flowing between them and the amount of electron or hole current flowing between the quantum dot and the third electrode have a non-linear relationship.

In the learning of artificial neural networks, the bond strength can be specified efficiently.

A circuit diagram of an example of a QD artificial neuron according to Example 1. The circuit diagram of another example of the QD artificial neuron according to Example 1. The schematic diagram explaining the solution principle concerning Example 1. FIG. The graph which explains the solution principle which concerns on Example 1. FIG. The graph which explains the effect which concerns on Example 1. FIG. The graph which explains the effect which concerns on Example 1. FIG. A circuit diagram of an example of a QD artificial neuron according to Example 2. The circuit diagram of another example of the QD artificial neuron according to Example 2. The graph which explains the solution principle which concerns on Example 2. A circuit diagram of an example of a QD artificial neuron according to Example 3. The schematic diagram of the artificial neuron according to Example 3. A circuit diagram of an example of a QD artificial neuron according to Example 3. A circuit diagram of an example of a QD artificial neuron according to Example 3. A circuit diagram of an example of a QD artificial neuron according to Example 3. The flow diagram explaining the flow of the learning process of the artificial neuron which concerns on Example 3. FIG. The graph which shows the relationship between the input and the output which concerns on Example 4. FIG. The graph which shows the effect which concerns on Example 4. FIG. A circuit diagram of an example of a QD artificial neuron according to Example 5.

Issues, configurations, and effects other than those described above will be clarified by the description of the following embodiments. Hereinafter, embodiments will be described in detail with reference to the drawings. However, the present invention is not limited to the description of the embodiments shown below. It is easily understood by those skilled in the art that a specific configuration thereof can be changed without departing from the idea or purpose of the present invention.

In the configuration of the invention described below, the same reference numerals may be used in common among different drawings for the same parts or parts having similar functions, and duplicate description may be omitted.

When there are a plurality of elements having the same or similar functions, they may be described by adding different subscripts to the same reference numerals. However, if it is not necessary to distinguish between multiple elements, the subscript may be omitted for explanation.

Notations such as "first", "second", and "third" in the present specification and the like are attached to identify components, and do not necessarily limit the number, order, or contents thereof. is not it. Further, the numbers for identifying the components are used for each context, and the numbers used in one context do not always indicate the same composition in the other contexts. Further, it does not prevent the component identified by a certain number from having the function of the component identified by another number.

The position, size, shape, range, etc. of each configuration shown in the drawings and the like may not represent the actual position, size, shape, range, etc. in order to facilitate understanding of the invention. Therefore, the present invention is not necessarily limited to the position, size, shape, range and the like disclosed in the drawings and the like.

The inventor of the present application creates a special artificial neuron element that spontaneously gives a bond strength when an input and an output are given, and if a neural network is constructed by this, the bond strength can be obtained without searching by trial and error. I thought it might be. For this concept, we devised a special artificial neuron in which the bond strength spontaneously becomes an appropriate size under the input and output given according to the natural force to become energetically stable. The conventional general artificial neuron element is a method of determining the output depending on the input and the bond strength, and its function and usage are different from those of the artificial neuron element devised by the inventor.

In the following examples, we will explain how to create an artificial neuron element that spontaneously optimizes the bond strength when inputs and outputs are given, and how to network them. In the embodiment, the non-linear electrical conduction of quantum dots (QD) is used. Quantum dots are microstructures of semiconductors or metals typically tens of nanometers in size. Quantum dots have non-ohm resistance in which current is not proportional to voltage, whereas ordinary conductors have ohm resistance in which voltage and current are proportional.

By using a structure or configuration that combines quantum dots, capacitance, electrical resistance, and power supply (hereinafter referred to as QD artificial neuron), the input / output relationship can be designed so that the relationship between the output current and the input current is a non-linear function. .. When the input / output relationship deviates from this nonlinear function, the charging energy of the capacitance is unstable, so the energy stabilizing action changes the input / output relationship so that it rides on the designed nonlinear function. At this time, a flow is spontaneously generated in which a current is discharged to the outside through a resistor, and conversely, a current is supplied from the outside. Since this current can be associated with the bond strength of the artificial neuron, the optimum bond strength according to the input and output can be known by measuring the current.

A neural network having a QD artificial neuron as a component constitutes a circuit provided with a current source, an ammeter, and a structure or element having a current mirror effect (for example, Non-Patent Document 3). The current flowing in the circuit when the current is passed through the current source can be measured, and the connection strength of the network can be obtained from it.

In Example 1, the composition, principle, and result of the smallest-scale artificial neuron using quantum dots are described. First, the configuration of the first embodiment will be described.

FIG. 1 is a diagram showing an example of a basic circuit configuration of a QD artificial neuron. The components of the QD artificial neuron are quantum dots 001, electrical resistance 002, capacitance 003, potential 004, ammeter 008,009, and current source 010, 011. The quantum dot 001 is realized by a known method such as Non-Patent Document 1. The circuit configuration as shown in FIG. 1, including the quantum dots 001, can be created by applying a semiconductor process.

The characteristics of quantum dots are that they have a nanoscale spatial region, that they can enter and exit the spatial region by tunnel effect, and that they can control the ingress and egress of electrons between quantum dots and the outside. Various electrical characteristics can be designed by utilizing this feature. Quantum dots are made, for example, by using compound semiconductors.

For example, in a laminated structure of two types of compound semiconductor layers of AlGaAs and GaAs, when Si is doped in AlGaAs, a layer in which conduction electrons called two-dimensional electron gas are accumulated is formed at the boundary between AlGaAs and GaAs. This electron can conduct in the x and y planes, but not in the z direction. That is, the electrons are confined in the z direction. Further, by applying a negative voltage to the gate electrode provided on the surface of the AlGaAs / GaAs laminated structure, a wall of electrostatic potential can be formed in the vicinity of the gate electrode in the two-dimensional electron gas layer surface. This enables confinement in the x and y directions. The cavity of the electrostatic potential in which the electron is confined corresponds to the quantum dot. The region where electrons exist outside the quantum dot acts as an electrical terminal for the quantum dot.

Capacitance 003 is a design matter and may include parasitic capacitance. Similarly, resistance 002 is also a design matter. Electrical resistance is assumed not only for integrated components such as chip resistors, but also for resistance derived from structures and shapes such as quantum wires and quantum point contacts by known semiconductor processes. Further, the electrodes 005, 006, 007 are connected to the current source 010, the current source 011 and the potential 004, respectively. The electrodes 005,006,007 may be, for example, a two-dimensional electron gas spreading outside the above-mentioned quantum dots, and may or may not have a physical electrode structure.

FIG. 2 is a diagram showing another circuit configuration example of a QD artificial neuron. Instead of the ammeters 008 and 009 in FIG. 1, the current is indirectly measured by the ammeter 012, 013, which is a combination of an electric resistance and a voltmeter connected in parallel.

The potential 004 merely indicates that the potential of the electrode 007 is kept constant with respect to time, and does not necessarily indicate the close relationship between the voltage source and the quantum dots. The current sources 010, 011 and ammeters 008,009 and voltmeters merely illustrate the connection relationship and do not always exist in the vicinity of the quantum dot 001.

The grounds 019-1 to 019-3, 1900 to which the current source 010, the current source 011 and the electric resistance 002 and the capacitance 003 are connected mean the reference potentials, and the potentials of the present embodiment are not necessarily the potentials of the earth. It is within expectations. Moreover, each may have a different potential. It is within the assumption of this embodiment that the current sources 010 and 011 are used not only in a constant manner with respect to time but also in a variable manner. Further, since this embodiment describes only the minimum configuration requirements, when circuit elements such as electric resistance, capacitance, inductance, ammeter, and voltmeter are added to the circuit of FIG. 1 or 2. It is also within the assumption of this embodiment that the means for estimating the current is taken by measuring the electric potential or the electric potential.

In the configurations shown in FIGS. 1 and 2, all of the illustrated configurations may be configured as a semiconductor device, for example, by one semiconductor chip. Alternatively, the quantum dots 001, electrodes 005 to 007, a part or all of the resistance and capacitance are composed of a semiconductor device, and the current sources 010, 011 and the ammeters 008, 009, 012, 013 are used as separate devices outside the semiconductor device. , May be measured via a terminal provided in the semiconductor device.

Next, we describe the principle that the movement and connection strength of QD artificial neurons are specified. As shown in FIGS. 1 and 2, the quantum dot 001 has three electrodes 005,006,007. When the inventor appropriately adjusts the thickness of the tunnel barrier of the quantum dots and the voltage of the power supply, the charging energy U of the capacitance 003 is the current i _in and the output current I _out (= i _out ) in FIGS. 1 and 2. We have found that it is possible to design an energy valley structure that minimizes under certain conditions.

As an example of the adjustment method, in FIG. 1, the thickness of the tunnel barrier between the quantum dot 001 and the electrode 007 is made sufficiently thinner than the thickness of the tunnel barrier between the quantum dot 001 and the electrode 005 and the electrode 006. Further, the potential 004 is sufficiently negative or positive as compared with the voltage of the electrodes 005 and 006 (the sign is inverted depending on whether the electron or hole is confined in the quantum dot).

By making adjustments, in the circuit configuration of FIG. 1, electrons or holes are stably flowed from the potential 004 to the electrodes 005 and 006 via the quantum dots 001, and _{the relationship between the current i in} and the current i _out becomes a non-linear function. be able to. An example of a nonlinear function is the relationship shown in FIG. 4, which will be shown later. The nonlinear function realized by the quantum dot can be associated with the activation function in the node of the artificial neuron.

FIG. 3 is a conceptual diagram showing the valley structure of the energy. According to the research of the inventor, the condition for the minimum energy (valley bottom 301) in FIG. 3 is when the current i _in and the output current I _out (= i _out ) in FIG. 1 or 2 satisfy Equation 1.

Here, e = + -1.602 × 10 ^-19 Coulomb (positive or negative depends on the electron and hole) is the elementary charge, and Γ is the barrier between the quantum dot 001 and the electrode 005,006. The average number of times an electron or hole can tunnel through a barrier per second. As an example of the adjustment method, when the thickness of the tunnel barrier between the quantum dot 001 and the electrode 007 is sufficiently thinner than the thickness of the tunnel barrier between the quantum dot 001 and the electrode 005 and the electrode 006, the quantum is quantum. The average number of times an electric charge can be tunneled between the dot 001 and the electrode 007 in one second is about 100 times or more the above-mentioned Γ.

_{By the way, since I in} = i _in + di _f from the conservation law of electric current, Equation 1 can also be written as Equation 2.

FIG. 4 is a diagram _{for explaining the process of reaching an energetically stable state when the input current I in} (i _in ) and the output current I _out of the QD artificial neuron are determined, and the curve 401 in the figure is shown. Corresponds to the valley bottom 301 when the energy valley structure of FIG. 3 is viewed from above. The black circles represent the state of t = 0, and the white circles represent the state of t >> 0. Assuming that time t = 0 and di _f = 0, i _in = I _in . In general, i _in = I _in does not satisfy the stability condition shown in Equation 1, so that the state is energetically unstable. Therefore, when _{the current of di f} flows, the QD artificial neuron voluntarily adjusts _{i in so as to satisfy Equation 1.}

By the way, the input / output relationship of an artificial neuron is generally described by Equation 3.

Here, x is an input, y is an output, w is a bond strength, and h is a nonlinear function. Equation 3 can also be rewritten as y = h (x + dx) by dx = (w-1) x. Therefore, Equation 2 is obtained by rereading the equation 3 x → _I in, and dx → -di _f. As described above, the QD artificial neuron functions as an artificial neuron, and its connection strength w can be known from _{the current di f.}

With reference to FIG. 5, the result of confirming the above principle by simulation is described. FIG. 5 is a plot of i _{in −} I _out graphs measured after a certain period of time (i _in , I _{out ) when various input currents I in} and I _out were applied to the QD artificial neuron. .. It can be confirmed that all the points are on the theoretical line shown by Equation 1 predicted by the principle. This result is the opposite, which means also have a function of di _f spontaneously adjusted to ride on (i _{in, I out)} is of the formula 1 conditions.

In the simulation of FIG. 5, in order to simplify the explanation, it is assumed that the value R _f of the resistor 002 in FIGS. 1 and 2 is sufficiently small.

FIG. 6 shows the simulation results when _{R f is sufficiently increased.} In Figure _6, Γ ⁼ 10 8 Hz, has a simulation under the condition of R f = 10 ⁶ Ω. The difference from the simulation in FIG. 5 is that the conditions for stabilization (i _in , I _out ) depend on the input current I _in. According to the simulation, the result is that it shifts _in the i- _{in axis direction as a whole depending on I-in.} Based on the result of the simulation, Equation 1 is modified as Equation 4.

k and δ _sft are parameters that depend on _{R f.} The quantum dots used in this embodiment shall be realized by a conventionally known method. As a realization method, for example, various methods such as those using the semiconductor laminated structure described above and those by self-formation described in Non-Patent Document 1 are known. In this embodiment, the method of realizing the quantum dots is not particularly limited. Moreover, in the above, the carrier is assumed to be a hole as well as an electron. FIGS. 3 to 6 define and display the direction in which electrons move as a positive current flow.

In Example 2, another configuration of the QD artificial neuron is described. 7 and 8 are partially rearranged versions of FIGS. 1 and 2, respectively.

FIG. 7 shows that the ammeter 008 is arranged on the output side and the current _dib is measured as compared with the configuration of FIG.

In FIG. 8, as compared with the configuration of FIG. 2, a combination 012 in which an electric resistance and a voltmeter are connected in parallel is arranged on the output side, and the current _dib is measured.

FIG. 9 is a diagram for explaining the process of reaching an energetically stable state when the _{input current i in} and the output current I _out (i _out ) of the QD artificial neuron are determined in the cases of FIGS. 7 and 8. The curve 901 in the figure corresponds to the valley bottom of energy. The black circles represent the state of t = 0, and the white circles represent the state of t >> 0. Also in the configuration of FIGS. 7 and 8, as shown in FIG. 9, adjustment is made voluntarily current di _b, reaching a stable state. At this time, the stable current state can be described as shown in Equation 5.

Adjustment is made voluntarily current di _b, in that reaching a stable state is the same as in Example 1, it can be used as an artificial neuron.

In Example 3, an example of a neural network using the QD artificial neuron described in Example 1 and Example 2 will be described.

FIG. 10 shows an example of a neural network having a QD artificial neuron as a component. The current sources 101 to 104 represent current sources that are constant or fluctuate with respect to time. Resistors 002, 105-108, 113-116 are not only integrated parts such as chip resistors, but also narrow constrictions sandwiched between so-called quantum wires and conductors that constrain the movement of electrons in one-dimensional direction in a solid. Resistance derived from the structure and shape such as so-called quantum point contact whose width is about the same as the electron wavelength is also assumed. The ammeters 109 to 112 and 117 to 120 include those configured by a combination of a resistor and a voltmeter as shown in FIG. The arrangement of the ammeter is an example and is not limited to this configuration, but it is desirable to obtain the current value flowing through the resistor 002.

The part surrounded by the dotted line represents a current mirror element (including the concept of a current mirror structure having an equivalent function thereof; the same applies hereinafter) 121-1 and 121-2. This is an element (or structure) characterized in that when a current flows through one side, the same amount of current flows through the other side. That is, electrons or holes 1000 flow symmetrically in a symmetrically configured circuit. Non-Patent Document 3 is an example of an experiment.

The shaded blocks show the QD artificial neurons 123 to 126, and the part excluding the current sources 010,011, ammeter 008,009, and electrical resistance 002 in FIG. 1 is displayed as a black box. The QD artificial neurons 123 to 126 are connected to the reference potential 1900 via the electric resistance 002 as shown in FIG. 1 and the like. The plurality of reference potentials 1900 do not necessarily have to be the same potential.

In FIG. 10, the reason why the current mirror elements 121-1 and 121-2 are arranged between the QD artificial neurons 123, 125 and 124, 126 is as follows. In the QD artificial neuron shown in FIG. 1, the electrons or holes that have flowed from the electrode 007 connected to the potential 004 to the quantum dots 001 are configured to escape to the electrodes 005,006. Therefore, the directions of the currents are opposite on the input side and the output side of the quantum dot 001. Therefore, in order to avoid complicated configuration when networking QD artificial neurons in multiple stages, the direction of the current is reversed by the current mirror element 121, and the current source 010 is unidirectionally directed to the current source 011. The current is allowed to flow.

FIG. 11 shows a schematic diagram of a generally known two-layer neural network. As is well known, in a neural network, when there are a plurality of inputs of the artificial neuron 1100, the sum of the products of each input and the connection strength is the input x of the artificial neuron, and the nonlinear function σ (x) is the output.

The configuration of FIG. 10 can correspond to the neural network of FIG. The number and connection relationships of artificial neurons 1100 in neural networks vary widely, and FIGS. 10 and 11 are just examples. In the configurations shown in FIGS. 10 and 11, the number of inputs can be increased by increasing the number of parallels. In addition, the number of layers can be increased by increasing the number of series. All the neural networks having the same parallel and serial connection relations as in FIG. 10 are within the assumption of the third embodiment. Finally, an example is given to show that machine learning is possible with a neural network whose component is a QD artificial neuron.

FIG. 12 is a 2-input 6-layer neural network having eight QD artificial neurons 203-1 to 203-8 as components. A QD artificial neuron 203-1 to 203-8 is arranged between the current source 201 and the current source 202. The QD artificial neuron 203 corresponds to the portion of FIG. 1 excluding the current sources 010, 011 and ammeter 009 and resistor 002. Each QD artificial neuron 203 is connected to the reference potential 1900 via the electrical resistance 002, as in FIGS. 1 and 10. A current mirror element 121 is arranged between the QD artificial neurons 203 in each layer.

The machine learning flow begins by passing current through _{the input current i in} and the output current I _{out of this neural network.} It _{is assumed that i in} and I _out fluctuate with time, but for the sake of simplicity, I _in = -15pA and I _out = 5pA. When the current starts to flow in this way, the current also starts to flow in each path in the network. Such a current amount is finally stabilized under the condition that the charging energy of each QD artificial neuron is minimized, as described with reference to FIGS. 4 and 9.

FIG. 13 shows the amount of current in each path after a sufficient time has elapsed in the circuit shown in FIG. How long it takes to reach a stable state depends on the design of electrical resistance and capacitance, but it is possible to reach a stable state even in 1 microsecond, and depending on the design, it will reach a stable state in a shorter time. It is also possible to reach. The currents of each path when the stable state is reached are shown in association with the arrows in FIG. The numbers attached to the arrows in FIG. 13 indicate the amount of current, and the arrows indicate the positive direction of the current.

Finally, record each amount of current. A particularly important amount of current is the amount of current of di _{f1 to di f8} in FIG. The currents of di _{f1 to di f8} are the currents flowing between the QD artificial neuron and the reference potential 1900. Each reference potential 1900 does not necessarily have to be the same potential. That's all for machine learning. When the neural network is operating, the recorded _{currents of di f1 to di f8} are set to obtain a desired output for a desired input.

Finally, in order to confirm that machine learning is possible, when the current value is set in the model of the neural network in FIG. 12 and I _in = -15pA is input, the desired I _out = 5pA is obtained. Make sure. The model of the neural network in FIG. 12 is specifically a mathematical model represented by equations 6 to 13. As can be seen from Equations 6 to 13, each QD artificial neuron corresponds to a node of the neural network, and the output of the node with respect to the input is determined by the _{current values di f1 to di f8.} Therefore, it can be said that the coupling strength (weight) between the nodes _{corresponds to the current values di f1 to di f8.}

here,

And δ _i is the correction coefficient. In this case, k = 3, η = 0.68, δ _i = 0.08pA. Calculating Equation 6 Equation 13 a chain _{reaction, I out} is determined from _{I in.} At this time, the amount of current measured by the neural network of the QD artificial neuron shown in FIG. 13 is set in the _{variables di f1 to di f8.}

When I actually _calculated, at the time of I in = -15pA, for the desired _I out = _5pA, obtained I out = 4.91pA, had been able to machine learning in the neural network of the QD artificial neuron Can be confirmed.

FIG. 14 shows the amount of current of each path calculated based on the equations 6 to 13. Other current values are the same as in FIG. Comparing FIG. 14 with FIG. 13, not only the final output but also the intermediate current can be reproduced well.

The most important point in the above machine learning is that in existing machine learning such as backpropagation, the optimum value of the bond strength must be searched by trial and error, but the machine learning of this embodiment is based on trial and error. The point is not to search for the optimum location, but simply to wait until it becomes energetically stable. The problem that machine learning increases exponentially with respect to the scale of the neural network is caused by a search by trial and error. Therefore, this can be avoided by using the machine learning of this embodiment, and the quality of machine learning can be avoided. Expected to speed up.

FIG. 15 summarizes the above flow of machine learning. The learning of the artificial neural network according to this embodiment is performed in two steps. The first step 1501 is a process by a device using the QD artificial neuron shown in FIGS. 10 and 12. In the process of the first step 1501, first, in the QD artificial neuron, desired input data and output data are _{set as current values I in} and I _out given to the QD artificial neuron (S1502). In general, the input data corresponds to the problem of teacher data, and the output data corresponds to the answer of teacher data.

Next, after the current value flowing through each part of the QD artificial neuron is stabilized, the current value _dif or _dib is read (S1503). The current value _dif or _dib is recorded separately as a parameter set in the artificial neural network. The first stage can be replaced by a simulator using a computer, as will be described later.

The second step 1504 is a process for an artificial neural network programmed to process the problem. The artificial neural network can be configured in software by imitating the circuits shown in FIGS. 10 and 12. Further, it is known that the artificial neural network shown in FIG. 11 can be implemented in hardware such as FPGA (Field-Programmable Gate Array). The specific configuration of the artificial neural network may follow the known technique.

In the second stage 1504, the coupling strength is learned by setting _{the current value dif} or _dib recorded in the first stage 1501 as the parameters of the artificial neural network configured by software or hardware.

In the explanation of the learning of the first embodiment, _{the processing when the current values I in} and I _out are set to be constant has been described as an example for explaining the principle. However, it is common that there are multiple teacher data to be trained by the neural network. If there is a large amount of teacher data, it can be dealt with by changing the amount of _{I in} and I _{out currents over time.} Specifically, for example, the current sources 101 to 104 in FIG. 10 are used as current sources that fluctuate with respect to time, and _{the current amounts of a plurality of I in} and I _out corresponding to the problem and answer of the teacher data are applied at regular time intervals. Switch to.

When operated in this way, the charging energy of the QD artificial neuron changes with time. An example of such modulation of energy (potential) with respect to time is an ion trap. In an ion trap, it is known that ions move according to the average potential with respect to the temporally modulated potential, and if there is a dent in the potential, it is trapped there. Similarly, in the case of QD artificial neurons, if the input and output are repeatedly modulated, the charging energy should behave so that its average is minimized.

Therefore, if there are a large number of teacher data _{, the current amount of I in} and I _out of a large number of sets is switched at regular intervals, and after a predetermined time, _{the values of di f1 to di f8} stabilize, and the learning is completed. good.

FIG. 16 is a conceptual diagram showing an example _in which I _{in and I out} are changed over time in order to handle a plurality of teacher data. This example deals with a simple model in which the answer I _out is constant, the problem I _in is modulated from 0 to H, and this is repeated multiple times. For I _in modulation, for example, the current source 201 in FIG. 12 is modulated.

FIG. 17 is a conceptual diagram showing the results of simulating the initial state and the state at the time when a sufficient time has passed when the input / output of FIG. 16 is given. The horizontal axis is i _in and the vertical axis is I _out.
The stability condition is shown by line 1700. The direction of the curve is different, but this is similar to the graphs of FIGS. 4 to 6. The plurality of circles in the graph indicate the current values of the QD artificial neurons in the circuit (di _{f1 to di f8 in} FIG. 12).

In the initial state (a), the current value of the QD artificial neuron does not converge to the stable line 1700. However, it can be seen that "learning" progresses with the passage of time, and the current value measured after a sufficient standby time converges to the line 1700, which is a stable condition.

Therefore, if there are a plurality of teacher data _{, the modulation of at least one of the input I in} and the output I _out may be repeated a plurality of times, and the current value may be measured after a predetermined time elapses.

As shown in FIG. 10, in the third embodiment, the QD artificial neurons are combined by arranging the current mirror element 121, and the carrier flows to the ground potential of the current mirror element 121, for example. On the other hand, in the fifth embodiment, an example in which the current mirror element is not arranged will be described.

FIG. 18 is a diagram showing a basic structure of the fifth embodiment. As shown in FIG. 18, the neural network is configured by combining a circuit unit 1601 having a hole (or electron) 1001 as a carrier and a circuit unit 1602 having an electron (or hole) 1002 as a carrier. As shown in FIG. 18, since the electrons 1002 and the hole 1001 propagate or conduct in opposite directions to each other, it is possible to create the same carrier movement as the current mirror element 121 shown in FIG. The holes 1001 and the electrons 1002 disappear by recombination at the boundary between the circuit unit 1601 and the circuit unit 1602.

Note that FIG. 18 is merely an example of a neural network configured by combining a circuit unit 1601 having a hole as a carrier and a circuit unit 1602 having an electron as a carrier. How to configure the circuit unit 1601 having a hole as a carrier and the circuit unit 1602 having an electron as a carrier is only a design matter.

The behavior of the device or circuit of Examples 1 to 4 can also be simulated. That is, instead of the electronic circuit shown in FIG. 12, a simulator that simulates the operation of the electronic circuit on a computer can be used, and the calculation of the amount of current can be performed on the simulator. Even in such a case, since the same effect can be expected based on the idea of the above-mentioned embodiment, it is included as one of the examples.

Machine learning on a conventional computer requires searching by repeating trials and errors until the desired output is obtained, adjusting the bond strengths that are independent of each other one by one, which is inefficient and coupling. There was a problem that the calculation time increased exponentially with respect to the number of intensities. On the other hand, according to the embodiment described in detail above, learning can be performed in a shorter time by using the QD neural network. That is, in the QD neural network, the adjustment of the excess and deficiency of the current at the connection point between the QD artificial neurons, which corresponds to the optimization of the connection strength, is spontaneously and simultaneously performed in all the QD artificial neurons, so the connection strength is tried. It can be identified efficiently without error.

001 ... Quantum dot 002 ... Electric resistance 003 ... Capacitance 004 ... Potential 005,006,007 ... Electrodes 008,009 ... Ammeter 010,011 ... Current source 012 , 013 ... Ammeter 101, 102, 103, 104 ... Current source 121 ... Current mirror structure or element

Claims (15)

Equipped with quantum dots, capacitance section, electrical resistance section,
The quantum dots include a first electrode, a second electrode, and a third electrode.
The first electrode is connected to the first potential and
The second electrode is connected to the first current source.
The third electrode is connected to a second current source.
The amount of current of electrons or holes flowing through the electric resistance section can be measured.
Electronic circuit. Electrons or holes are stably flowed from the first potential to the first electrode and the second electrode via the quantum dots.
The amount of electron or hole current flowing between the quantum dot and the second electrode and the amount of electron or hole current flowing between the quantum dot and the third electrode have a non-linear relationship.
The electronic circuit according to claim 1. The tunnel rate between the quantum dot and the first electrode is larger than the tunnel rate between the quantum dot and the second electrode and the tunnel rate between the quantum dot and the third electrode.
The electronic circuit according to claim 1. The capacitance section and the electrical resistance section are arranged in parallel with the path between the second electrode and the first current source.
The electronic circuit according to claim 1. The capacitance section is arranged in parallel with the path between the second electrode and the first current source, and in parallel with the path between the third electrode and the second current source. The electric resistance part is arranged,
The electronic circuit according to claim 1. Further provided with a second quantum dot, a second capacitance section, and a second electrical resistance section.
The second quantum dot comprises a third electrode, a fourth electrode, and a fifth electrode.
The third electrode is connected to the second potential and
The fourth electrode is connected to the third electrode, and the fourth electrode is connected to the third electrode.
The fifth electrode is connected to the second current source.
A current mirror element was arranged between the fourth electrode and the third electrode to form a network configuration.
The electronic circuit according to claim 1. Further provided with a second quantum dot, a second capacitance section, and a second electrical resistance section.
The second quantum dot comprises a third electrode, a fourth electrode, and a fifth electrode.
The third electrode is connected to the second potential and
The fourth electrode is connected to the third electrode, and the fourth electrode is connected to the third electrode.
The fifth electrode is connected to the second current source.
The carriers flowing through the quantum dots, the capacitance section, and the electric resistance section are electrons (or holes), and the carriers flowing through the second quantum dots, the second capacitance section, and the second electrical resistance section are holes (holes). Or electronic), with a network configuration,
The electronic circuit according to claim 1. Modulates at least one of the first current source and the second current source in time.
The electronic circuit according to claim 1. Equipped with quantum dots, capacitance section, electrical resistance section,
The quantum dots include a first electrode, a second electrode, and a third electrode.
The first electrode is connected to the first potential and
The amount of current of electrons or holes flowing through the electric resistance section can be measured.
Electronic circuit,
It is a learning method of a neural network using
The first step of setting the current value corresponding to the problem of the teacher data to the first current source connected to the second electrode.
The second step, in which the current value corresponding to the answer of the teacher data is set in the second current source connected to the third electrode.
The third step of measuring the amount of current of electrons or holes flowing in the electric resistance portion,
Fourth step of constructing a neural network corresponding to the electronic circuit,
Fifth step of setting a value corresponding to the amount of current as the coupling strength of the neural network,
How to learn neural networks including. Prepare multiple electronic circuits
By connecting the second electrode and the third electrode, a multi-stage configuration is formed.
The first current source is connected to the second electrode of the first stage electronic circuit.
The second current source is connected to the third electrode of the final stage electronic circuit.
A current mirror element is arranged between the second electrode and the third electrode.
In the third step, the amount of current flowing through the electric resistance portion of the electronic circuit constituting each stage is measured.
In the fourth step, a multi-layer neural network is configured as the neural network.
In the fifth step, the coupling strength of each layer of the multi-layer neural network is set to a value corresponding to the amount of current flowing through the electric resistance portion of the electronic circuit constituting each stage.
The method for learning a neural network according to claim 9. Prepare multiple electronic circuits
By connecting the second electrode and the third electrode, a multi-stage configuration is formed.
The first current source is connected to the second electrode of the first stage electronic circuit.
The second current source is connected to the third electrode of the final stage electronic circuit.
The carrier of the electronic circuit having the multi-stage configuration has a configuration in which electrons and holes are alternately used.
In the third step, the amount of current flowing through the electric resistance portion of the electronic circuit constituting each stage is measured.
In the fourth step, a multi-layer neural network is configured as the neural network.
In the fifth step, the coupling strength of each layer of the multi-layer neural network is set to a value corresponding to the amount of current flowing through the electric resistance portion of the electronic circuit constituting each stage.
The method for learning a neural network according to claim 9. The electric resistance portion has a path between the second electrode and the first current source and a predetermined potential, and a path and a predetermined potential between the third electrode and the second current source. Placed in at least one in between,
The method for learning a neural network according to claim 9. Instead of the electronic circuit, a simulator that simulates the operation of the electronic circuit on a computer is used, and the first to third steps are performed on the simulator.
The method for learning a neural network according to claim 9. The tunnel rate between the quantum dot and the first electrode is larger than the tunnel rate between the quantum dot and the second electrode and the tunnel rate between the quantum dot and the third electrode.
The method for learning a neural network according to claim 9. When using a set of multiple questions and answers as the teacher data,
The current value set in the first current source is modulated, the modulation is repeated a plurality of times, and after a predetermined time, the third step is performed.
The method for learning a neural network according to claim 9.

Images