JPH0348364A - Matrix/vector product sum computing element - Google Patents

Abstract

PURPOSE:To improve the performance of a system by using a constitutional part consisting of a magnetic thin film capable of holding magnetic bubbles, two conductor lines intersecting with each other on the thin film and a magnetic detector arranged on the thin film and connected with the conductor line as a unit. CONSTITUTION:A bubble holding garnet film 2 is a ferrimagnetic thin film formed on a Gd3Ga5O12(111) substrate 1 by epitaxial method having the film face perpendicular direction as the easy axis of magnetization. Conductor lines 4 arrayed in the vertical direction are X leads, conductor lines 5 arrayed in the horizontal direction are Y leads and both the X and Y leads are allowed to intersect with each other on respective cells. The detector 6 utilizing the magneto-resistance effect is shown on the upper right area of the matrix and the detectors 6 of respective cells are connected in series in the Y direction similarly to the Y leads as shown by 7. Consequently, the calculation speed can be improved, a matrix storage part can be formed as a completely non- volatile part and the system performance can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a matrix/vector product-sum calculation element incorporating a nonvolatile high-density solid-state magnetic storage element.

(Conventional technology) The basic technology is provided by Electronics (Ohmsha Co., Ltd., 1)
988. August issue, pages 52-56, written by Hiroo Yonezu) and Hideki Aso, Binal Network Information Processing (Sangyo Tosho Co., Ltd., published June 1988, pages 20-27)
Explain by quoting. In engineering models of neural networks, the outputs of one neuron are interconnected to the inputs of other neurons, just like in a real neural network. There are generally three types of connections called synaptic connections: 10, 1 and 0. + means that the output of a neuron is directly connected to the human power of another neuron (positive connection). One is to reverse the output of a neuron, make it one, and connect it to the input of another neuron (
negative bond). The former corresponds to excitatory connections, and the latter corresponds to inhibitory connections. O means not bonded. When the number of neurons is N, Figure 5 (
A coupling circuit like a) is created. This pattern of connections is called a connection matrix (memory matrix), and it is used to program the neural network. The output Vj of the jth neuron is the input ui of the ith neuron
The weights coupled to are written Wij and are called the coupling strength. Since the input ui entering the i-th neuron receives outputs from many other connected neurons, it is given by ui=ΣjWijVj+Ii. Here, Ii is an external input to the i-th neuron or a threshold value. The relationship between the input ui and the output Vi of a neuron can be expressed by a characteristic curve ``i''g(ui) as shown in Figure 5(b).This curve is a sigmoid (sigmoid)
moid) function (S-shaped function), when the human power becomes positive and large, it produces an output of 1, and when the input becomes a large negative value, it produces an output of 0.The limit where the rise of the sigmoid function becomes steep is the digital The coupling strength Wij, which determines the operation mechanism and the contents of memory, is the effective conductance.

I1 to I in the connected neural network of FIG.
When N inputs are input all at once, the circuit operates in conjunction with each other and often settles into a certain state. The output of each neuron at this time becomes the output for human power. The above is the model proposed by Hopfield.

What is distinctive about neural networks is that instead of just one CPU operating at high speed in time series as in conventional computers, N neurons are operating all at once in parallel. An example in which the circuit shown in FIG. 5 is applied to an associative memory will be described. FIG. 6 illustrates its contents. First, in order to memorize three types of living things, outputs (1) and (2) are considered to be information regarding the shapes of two-legged and four-legged creatures, respectively. Consider ■3 and ■4 as information regarding the athletic ability of flying and walking, respectively, and consider ■5 and ■6 as information regarding the size of height and shortness.

Therefore, assuming that signal 1 corresponds to the above information and signal 0 does not correspond, humans, dogs, and birds can be represented by 6-bit information as shown in FIG. For example, humans learn that they are tall because they walk on two legs.

The combined matnox (Wij) when these three 6-bit information are memorized is given by the following equation in the model by Hopfield.

Wij =Σ,=13(2Vi(8)-1X2Vj(a
)-1>Here, S means three types: humans, dogs, and crows. That is, three types of information are stored in an overlapping manner. here,
Partial information that lacks information regarding athletic ability such as having two legs and being tall (I1 = I5 = 1, I2 = I3: 1
4 = I5 = 0) as an input, the circuit of FIG. 5 operates while feeding back each other and settles into a certain state. The output at this time is two-legged, walking, and tall (Vl=V4=V5=1, V2=V3=V5=0)
If you get this, you have obtained complete information about humans.

This is a function called associative memory. Obtaining complete information from partial information in this way is similar to the situation that humans often encounter, where they see something for a moment and then go from having a vague memory to remembering it clearly after a while.

Therefore, this network is an important issue that is attracting attention in areas such as pattern recognition and speech recognition.

To give a more general explanation, Hopfield's network is an interconnected network, first proposed as a binary unit, and later as a multivalued (continuous, discrete) network.
Expansion to the unit was carried out. Here, we will introduce binary values for a basic understanding of the operation.

A feature of Hopfield's network as an interconnected network is that the connections between units are, in principle, symmetrical connections. That is, units i to j
If there is a connection with strength Wij to
There are connections Wji with the same strength.

In addition, state changes are asynchronous, and each unit detects inputs weighted by the strength of connections at arbitrary timing, sums them up, and changes the state based on that value to output an output. Writing this state change rule as a formula: If ΣjWjtuj+st ei>0, then the state of ui=1ΣjWjiuj+si− ei<o,
State of ui=0ΣjW34uj+si−ei=
o, the state of ui remains unchanged. Here, ΣjWjiuj is
The sum of inputs from other units, Si is the human input from outside the network, and Oi is the threshold of unit i. Asynchronous operation means that when one cell changes its output value, other cells maintain their output value.

In other words, this is an operating method in which two or more cells do not change their outputs at the same time. Each cell performs calculations independently of each other.

Here, an example of a simple self-remembering associative memory by Hopfield is shown. It is assumed that the network consists of 16 binary units. These units are arranged in a 4x4 square in Figure 4.

The state of the network becomes a 16-dimensional binary vector such as (1, 0.0..., 1). Appropriately three states α
, , α2, α3 and store them. That is, let them correspond to a stable equilibrium state of the network. To do this, we define the combination Wij as Wij=Σ,
(2ui(as)-1)(2uj(a,)-1),
Let WH = 0.

Starting from an appropriate initial state, the memorized pattern is recalled. What is new about Hopfield's network is that it makes the operation of the network easier to understand by limiting the connections to symmetric ones, introducing asynchronous operations, and introducing the concept of network energy.

Since the Hopfleld network includes a feeder loop, the equilibrium state of the network is important as a characteristic of its operation. Therefore, when using networks, it is important to investigate the properties of the equilibrium state.

Therefore, consider the following function of the state α of the network.

E (a) = ΣiΣj Wijuj (a) uj (a) ten Σ
i(si-ei)ui(a) However, u4(α) etc. are the values taken by unit i in state α. now,
Let's see how this function changes when a certain unit u4 changes state according to state change rule 6. If the value of ui changes from 1 to 0, the amount of change in E is ΔiE(α)=ΣiWjiuj(
Q) +Si − Oi.

By the way, according to the state change rule, the fact that the value of ui changed from 1 to 0 means that the sum of the inputs to ui is equal to the threshold value o, that is, ΣjWjHuj−sH−ei<0. However, although the left side and ΔiE above are equal, E has decreased due to the state change.

A similar argument can be made for the case where the value of ui changes from 0 to 1, and it can be seen that E decreases in either case.

E decreases each time a state change occurs, but no matter which unit's state changes in the network state, E decreases.
If the state is such that the value of increases, no further state changes occur. From this, it can be seen that an equilibrium state is a state in which the value of E takes a minimum value.

The equilibrium state can be characterized as giving a minimum value of the energy of the network, and the operation of the network can also be interpreted as minimizing its energy. This interpretation is possible because we limit bonds to symmetrical ones.

If you create and operate a network such that a certain evaluation function has just the energy, you can use that network to find a state that optimizes the evaluation function using parallel processing. In other words, networks can be viewed as optimization machines.

How to determine the connection Wij will be explained. In the above example, in order to store α1, α2, α3, Wij=Σ, :13(2ui(a,)-1X2uj(a
, )-1) Wii=0, si=0, ei=0.

Which one of the states you have now memorized by the network?
Suppose that the value is α8.

At this time, the total human power for unit i is ii=ΣiW
ijui(αs) =Σ, (2ui(a8)−1) {Σjuj(αs)
(2uj(αB) 1)) The values in {} vary depending on the value of S. So that this becomes 0 except when S=S,
If α1, α2, α3 are created well, ii=2ui(a,-1) The next state will also be 1. Also, if, ui(
If αs)=0, then ii<0, so the next state of ui will also be O. Therefore, the state of the network is
It can be seen that there is no change from α.

In terms of energy, this means that ΔE>O for all unit state changes. In other words, as is E
It gives the minimum of . Therefore, in order to successfully determine the connections described above, Σiuj(aB冫(2uj(a,) 1)=2Σju
j(α8)uj(α,)−Σiui(α,) is S=
It is preferable that the condition for S other than 8 to be 0 is satisfied (sufficient condition).

This condition is α1, α2, . A vector such that α3 is 1 for exactly half of the 16 dimensions, and furthermore,
This holds true if the number of 1's shared by two vectors is half of the number of 4's. However, like this α
, is selected, the above condition is also satisfied for a vector that is the inverse of that vector (1 and O are swapped), so such a vector will also be in an equilibrium state. In general, it is not so easy to bring only the things you want to remember into an equilibrium state.

Aside from this, it is important for this element to efficiently calculate the sum of products of the coupling matrix {Wij} and the vector (u,). {Wij} is given a priori.

(Problems to be Solved by the Invention) There is an increasing need for elements suitable for such calculation purposes, that is, hardware that can incorporate parallel processing as much as possible. To achieve this purpose, semiconductor memory elements (DR)
An example of applying AM) has been reported. Although DRAM exhibits high speed in digital storage, calculation of the product of a matrix and a vector as described above takes a surprising amount of CPU time. This takes time because the only way to do this is to calculate the product and sum of each element of the matrix and each component of the vector separately, and the product is calculated for each element sequentially, and then the sum is calculated. . This is because the DRAM is a digital device and cannot accumulate output voltages. In other words, each element has a ground potential, and from that level the voltage can only be determined by the externally applied voltage. Therefore, it is necessary to add up the voltages for each element again. An important technical challenge is to reduce the time spent on this part of the calculation. With conventional methods using semiconductor memory, it has been difficult to shorten calculation time due to the characteristics (structure) of the device itself. The present invention presents a new method to solve this problem.

(Means for Solving the Problems) In the present invention, a magnetic memory is used to eliminate the drawback that the calculation time cannot be shortened with a semiconductor memory.

The unit of the present invention is a component consisting of a magnetic thin film that can hold magnetic bubbles, two conductor wires that intersect on the thin film, and a magnetic detector that is also placed on the thin film and connected with a conductor. It is a matrix/vector product-sum calculation element.

(Example 1) An example in which a bubble memory is used instead of a semiconductor memory will be shown. Figure 1 shows an example of the structure of the element of the present invention, which is 9x9 for simplicity.
Consider an example consisting of a matrix of The bubble-retaining garnet film 2 is a ferrimagnetic thin film that is elongated to a certain extent on the Gd3Ga5012 (111) substrate 1 by liquid phase epitaxial method and whose axis of easy magnetization is perpendicular to the film surface. In order to prevent interference between the cells shown by 2 in the figure, each cell is separated.

Separation methods include completely scraping off the boundary area between cells as shown at 8 in Figure 2, or using ion implantation to weaken the magnetism in the area corresponding to the scraped area. Apply methods such as:

The conductor wire 4 running in the vertical direction in FIG. 1 is the X lead, and the conductor wire 5 running in the horizontal direction is the Y lead.

The X leads and Y leads cross each other on each cell. 6 shown in the upper right area of FIG. 1 is a detector that utilizes the magnetoresistive effect. The detectors of each cell are connected in series in the Y direction, as shown by 7, like the Y leads.

Circle 3 within the cell indicates a magnetic bubble. Next, the operation of this element will be explained. The preservation of the memory matrix is achieved by selectively writing magnetic bubbles into each cell arranged in the matrix. A magnetic bubble is created by applying each magnetic field from the X and Y conductor currents to the cell where the X and Y conductor lines intersect so that the composite bias magnetic field of the magnetic fields from the X conductor current and the Y conductor current is higher than the critical magnetic field for bubble generation. Selectively select the size of the image and write. As a matter of course, bubbles should not be generated only by the X or Y conductor current. In addition,
Another advantage of the magnetic memory is that in this device, the contents of the storage matrix are not changed very frequently, and therefore the time required to rewrite the written data is not required to be very fast.

What is required of this element is the speed of readout time. The example of FIG. 1 meets this requirement.

The reading method in the example of FIG. 1 will be explained. A current pulse corresponding to the magnitude (Oorl) of each force xk of the vector is applied simultaneously to each X conductor line running in the vertical direction. For example, a method is such that a current pulse is not applied to the conductor line j whose vector value is 0, and a current pulse of a predetermined amplitude is applied to the conductor line j' whose vector value is 1. Then, the detector output using the magnetoresistive effect of each cell connected in series in the horizontal direction (Y direction) will determine the state of each cell (whether there is a bubble or not) and the current pulse in the X conductor wire. A voltage change is generated based on the magnetoresistive effect, which corresponds to the product of whether or not the current is applied. The truth table is shown in FIG. Figure 3 (
5), the ij element of the matrix is ``1'' (
If a bubble exists), when the vector Xj is inserted by a certain amount, a voltage change corresponding to the output "1'" is generated on the detection line i. As shown in FIG. 3(b), the ij element is “0” (
If there is no bubble), even if "1'" is inserted into the vector j, the output of the detection line i will be (1011).As shown in FIG. ”(
Even if a bubble exists), the vector Xj is “0” to some extent
, only a voltage change corresponding to the output "0"' occurs on the detection line i. As shown in Fig. 3(d), when the U element is ゜“0′” (no bubble), if “0” is entered in the vector command j, the output of the detection line i will also be “0′”. Become. If a line connecting each of these detectors in series is connected to a constant current power source, the sum of voltage changes of each cell can be detected simultaneously for each row of the matrix. This sum corresponds to the product sum ΣjWijxj of the i-row element of the matrix and a certain portion of each vector.

The magnetoresistive effect due to magnetic bubbles within the cell is detected as follows. The difference between the ground state (state where no current pulse is applied to the X conductor wire) and the excited state (state where a current pulse is applied to the X conductor wire) of the detection line for each row is taken. The difference is as shown in Figure 4, when a current pulse is applied to the X conductor wire, the bubble diameter decreases from 9 to 10, and therefore the coupling between the magnetic bubble and the magnetoresistive detector becomes smaller, and the magnetoresistance due to the bubble magnetic field decreases. This occurs when the effect decreases.Now, (YSmLuCa
)3 (FeGe) 5012-based material, the mobility of the bubble domain domain wall will be about 5 m/see-Oe, so in order to reduce a bubble with a diameter of 5 pm to 311 m at the time of detection, see-Oe = 2X10 -7 need to be met. For example, the time width of the pulse bias magnetic field is set to 0.5p.
sec, a magnetic field amplitude of 0.40e is sufficient.

In other words, one product-sum calculation is completed in 0.5 psec.

In order to obtain the necessary information, it is necessary to repeat this calculation about 10 times, so the total calculation time is about 5 psec. This calculation time is 1000x using semiconductor DRAM.
When calculating the matrix of lO00, about 100 ps
This means that the processing speed is significantly higher than that required with ec.

Note that if the pulse bias magnetic field amplitude is increased by 10 times, the above calculation time will theoretically be further reduced by an order of magnitude.

(Example 2) There are many cases in which there are generally three ways of making this connection, which is called a synaptic connection: 10, 0, and 0. + means that the output of a neuron is directly connected to the input of another neuron (positive connection). One is to reverse the output of a neuron, make it one, and connect it to the human power of another neuron (
negative bond). The former corresponds to excitatory connections, and the latter corresponds to inhibitory connections. 0 means no bond. The simplest example of how to create these excitatory and inhibitory connections in an actual element is to prepare two memory matrices, one for excitatory connections and the other for excitatory connections. By using one for inhibitory binding, 3
It is possible to construct matrix/vector multiply-accumulate circuits with level connections. Storage matrices with cell structures of three or more levels can also be assembled in a similar manner.

(Effects of the Invention) According to the present invention, it is possible to improve the calculation speed of Hopfield's self-remembering associative memory, and to make the matrix storage part completely non-volatile.
System performance was significantly improved compared to the system using M.

[Brief explanation of drawings]

FIG. 1 is a diagram of the main parts of the device of the present invention. FIG. 2 is a diagram showing an example of a cell separation method. FIG. 3 is a diagram illustrating a product circuit of matrix elements and vector precepts in the present invention. FIG. 4 is a diagram illustrating the operation of the magnetoresistive detector of the present invention. FIG. 5 is a diagram showing a Hopfield model of a neural network. FIG. 6 is a diagram showing an example of the operation of an associative memory using the Hopfield model. 1: Magnetic bubble holding layer substrate, 2: Cell, 3: Magnetic bubble, 4: X conductor wire, 5: Y conductor wire, 6: Magnetoresistive detector, 7: Connection of each magnetoresistive detector, 8: Intercell separation zone, 9: magnetic bubble, 10: TJF air bubble.

Claims (1)

[Claims] A unit is a component consisting of a magnetic thin film that can hold magnetic bubbles, two conductor wires that intersect on the thin film, and a magnetic detector that is also placed on the thin film and connected with a conductor. Matrix/vector product-sum operation element.