TWI718677B - Data sensing device and data sensing method thereof - Google Patents

Abstract

A data sensing device and a data sensing method thereof are provided. The data sensing device includes a compensation signal generator, a weighting operator and an arithmetic operator. The compensation signal generator receives a basic input signal and a plurality of reference weighting values, and generates a compensation signal according to the basic input signal and the reference weighting values. The weighting operator has a plurality of memory cells, performs writing operation on the memory cells according to the weighting values based on an address information, and the weighting operator generates an output signal by the memory cells by receiving a plurality of input signals. The arithmetic operator performs an arithmetic operator on the output signal and the compensation signal to generate a compensated output signal.

Description

Data sensing device and data sensing method

The present invention relates to a data sensing device and a data sensing method thereof, and more particularly to a data sensing device and a data sensing method that can realize negative number input and is suitable for neural network-like operations.

With the evolution of electronic technology, artificial intelligence technology has gradually become a mainstream. And the similar neural network is an important way to perform artificial intelligence calculations.

Neural network-like networks need to be accompanied by a large number of operations. In the conventional technical field, digital circuits or high-end processors are often used to perform neural network-like operations. Such an approach requires relatively high hardware costs, and also requires relatively high power consumption to perform computing operations. Corresponding to this, the conventional technology uses non-volatile memory to perform neural network-like operations, which can effectively reduce the circuit area and reduce the required power consumption.

In the conventional technical field, the transconductance of a memory cell in a non-volatile memory can be used as a weight in a neural network. Through the product of the input signal received by the memory cell and the transduction value of the memory cell, a neural network-like arithmetic operation can be realized.

However, in practical applications, the arithmetic actions of similar neural networks often have negative input signals. In the hardware architecture, in order to realize an input signal with a negative value, the input signal needs to be expressed by subtracting two values, for example, the first value minus the second value. When calculating, the first value and the second value are successively As the input of the neural network, the second output is multiplied by the negative sign and then the first output is added. In this way, the neural network operation with negative numbers will increase the amount of calculation twice and improve the implementation. The difficulty of the above, and cause the increase of the circuit cost.

The invention provides a data sensing device and a data sensing method, which are used for realizing a neural network-like operation with a negative input signal.

The data sensing device of the present invention includes a compensation signal generator, a weighting operator and an arithmetic operator. The compensation signal generator receives a reference input signal and a plurality of reference weight values, and generates a compensation signal according to the reference input signal and the reference weight value. The weight calculator has a plurality of memory cells, the weight calculator writes the memory cells according to multiple weight values based on the address information, and the weight calculator allows the memory cells to receive a plurality of input signals to generate output signals. The arithmetic operation unit is coupled to the compensation signal generator and the weight operation unit, and operates the output signal and the compensation signal to generate a compensated output signal.

The data sensing method of the present invention includes: providing a compensation signal generator to receive a reference input signal and a plurality of reference weight values, and generating a compensation signal based on the reference input signal and the reference weight value; providing a weight calculator, based on address information, Perform write operations on multiple memory cells in the weight calculator based on multiple weight values, and make the memory cells receive multiple input signals to generate output signals; and provide an arithmetic calculator to perform calculations on the output signal and the compensation signal To generate the output signal after compensation.

Based on the above, the present invention uses a compensation signal generator to calculate a compensation signal, calculates an input signal offset to a positive signal to generate an output signal, and then generates an input that may have a negative signal by performing calculations on the output signal and the compensation signal. The signal is output after the signal is compensated. Under a simple hardware architecture, neural network-like operations with negative input signals are realized.

In order to make the above-mentioned features and advantages of the present invention more comprehensible, the following specific embodiments are described in detail in conjunction with the accompanying drawings.

Please refer to FIG. 1. FIG. 1 is a schematic diagram of a data sensing device according to an embodiment of the present invention. The data sensing device 100 includes a compensation signal generator 110, a weight operator 120 and an arithmetic operator 130. The compensation signal generator 110 receives a reference input signal V _base and a plurality of reference weight values WR. The compensation signal generator 110 generates the compensation signal CV according to the reference input signal V _base and the reference weight value WR. In detail, the compensation signal generator 110 may calculate the reference weight sum of the reference weight value WR, and _{multiply the reference weight sum by the reference input signal V base} to obtain the compensation signal CV. In other embodiments of the present invention, the compensation signal generator 110 may also be a data storage device, and the compensation signal CV may be calculated outside the compensation signal generator 110 and stored in the compensation signal generator 110.

On the other hand, in the embodiment of the present invention, the weight calculator 120 receives address information ADD and a plurality of weight values W ₁ ˜W _M. The weight calculator 120 may have a plurality of non-volatile memory cells, and the weight calculator 120 may perform write operations on the non-volatile memory cells _{based on the address information ADD according to the weight values W 1} ˜W _M. Among them, the transduction value of the non-volatile memory cell after the writing operation is set according to the weight value W ₁ ~W _M , for example, the weight value W ₁ ~W _M can be A times of the corresponding transduction value. , A can be any real number. In addition, the weighting calculator 120 further receives the input signals V ₁ ˜V _N , and causes the input signals V ₁ ˜V _{N to be} respectively applied to a plurality of non-volatile memory cells, and generates an output signal OUT accordingly. Please note here that when the neural network has a negative input signal, it can _{perform an offset operation based on the reference input signal V base} for all input signals, so that the offset input signal V ₁ ~V _{N is} all positive signals. In this way, the weight calculator 120 can _{effectively perform neural network-like operations according to the input signals V 1} to V _N , and generate the output signal OUT. Please note here that the sum of the weight values W ₁ ~W _M is equal to the sum of the reference weight values of the reference weight value WR.

On the other hand, the arithmetic operator 130 is coupled to the compensation signal generator 110 and the weighting operator 120, and performs arithmetic operations on the output signal OUT and the compensation signal CV to generate the compensated output signal COUT. The arithmetic operator 130 may subtract the compensation signal CV from the output signal OUT to generate the compensated output signal COUT. In this way, the data sensing device 100 can generate an output signal (that is, the compensated output signal COUT) that actually corresponds to the neural network-like operation performed on the input signal with a negative value.

Please refer to FIG. 2 below. FIG. 2 is a schematic diagram of an implementation of a data sensing device according to an embodiment of the present invention. The data sensing device 200 includes a compensation signal generator 210, a weight operator 220, and an arithmetic operator 230. In this embodiment, the compensation signal generator 210 receives the reference input signal V _base and the reference weight value WR, and _{multiplies the sum of the reference input signal V base} and the reference weight value WR to generate the compensation signal CV. In other embodiments of the present invention, the compensable signal generator 210 is a data storage element, and the compensation signal CV can be calculated outside the compensation signal generator 210 and stored in the compensation signal generator 210.

On the other hand, in the embodiment of the present invention, the weight calculator 220 has a plurality of memory cells MC1 ˜MCM. Based on the address information ADD, the weight calculator 220 performs a write operation on the memory cells MC1 to MCM according to a plurality of weight values W ₁ to W _M. In addition, the weight calculator 220 also makes the memory cells MC1 ˜MCM respectively receive a plurality of input signals V ₁ ˜V _N to generate an output signal OUT. Among them, the memory cells MC1 ˜MCM can be constructed on the same memory string 222. _{When the weight value W 1} ~ W _M is written to the memory cells MC1 ~ MCM _{, the transduction values of the memory cells corresponding to the weight values W 1} ~ W _M can be calculated first, and then the transduction values calculated according to the respectively calculated Guide value to perform write operations on memory cells MC1~MCM.

The weight operator 220 further includes a sense amplifier 221. The sense amplifier 221 is coupled to the memory string 222, and receives the read current IR generated by the _{memory string 222 according to the received input signal V 1} ~V _N. The sense amplifier 221 also compares the read current IR with a plurality of reference currents to generate an output signal OUT.

It is worth mentioning that the number of memory cells in the weight calculator 220 can be set according to the number of weight values W ₁ to W _M in the neural network, and there is no particular limitation. In addition, the memory cells MC1 to MCM in this embodiment may be flash memory cells.

The calculator 230 is coupled to the compensation signal generator 210 and the weight calculator 220. The arithmetic unit 230 receives the compensation signal CV generated by the compensation signal generator 210, receives the output signal OUT generated by the weight calculator 220, and performs arithmetic operations on the output signal OUT and the compensation signal CV to generate the compensated output signal COUT. The arithmetic unit 230 may include an adder 231, and the adder 231 may subtract the output signal OUT and the compensation signal CV to generate the compensated output signal COUT.

Please refer to FIG. 2 and FIG. 3 simultaneously. FIG. 3 is a flowchart of the operation of the embodiment of the present invention in FIG. 2. In step S310, the reference weight value WR may be stored in the compensation signal generator 210, and the sum of the reference weight values WR (reference weight sum) is calculated, and the product of the reference weight sum and the reference input signal is calculated to generate the compensation signal. In addition, in step S320, the input signals V ₁ ˜V _N are sent to the weight calculator 220, and the input signals V ₁ ˜V _N are respectively operated with corresponding weight values according to the address information to generate the output signal OUT. In step S330, the output signal OUT is subtracted from the compensation signal CV to generate the compensation output signal COUT.

Please refer to FIG. 4. FIG. 4 is a schematic diagram of another implementation of the data sensing device according to the embodiment of the present invention. The data sensing device 400 includes a compensation signal generator 410, a weight operator 420, and an arithmetic operator 430. The compensation signal generator 410 includes a memory string 411 formed by a plurality of reference memory cells MCR1 ˜MCR3. And compensation signal generator 410 based on the plurality of weights to the reference writing operation for the reference memory cell MCR1 ~ MCR3, and the reference memory cell MCR1 ~ MCR3 are provided a plurality of reference conductance value G ^{_'1, G'} ₂ and G _^'3 . In addition, the memory string 411 of the compensation signal generator 410 also receives the reference input signal V _base , and provides a read current through the bit line BL1 of the memory string 411 as the compensation signal CV.

On the other hand, the weight calculator 420 has a memory string 421 formed by a plurality of memory cells MC1 to MC3. The weight calculator 420 performs write operations on the memory cells MC1 to MC3 respectively according to the multiple weight values, so that the memory cells MC1 to MC3 provide multiple conduction values G ₁ , G _2, and G _{3, respectively} . It is worth mentioning that ^{the sum of the reference conduction values G′} ₁ , G ^′ ₂ and G ^′ _{3 is the same} as the sum of the conduction values G ₁ , G ₂ and G ₃ . In addition, the memory cells MC1 to MC3 respectively receive input signals V ₁ , V ₂ , and V ₃ , and provide a read current through the bit line BL2 of the memory string 421 as the output signal OUT.

The arithmetic operator 430 includes a sense amplifier (SA) 431 and a current subtractor 432. The current subtractor 432 receives the compensation signal CV and the output signal OUT, and subtracts the compensation signal CV from the output signal OUT (multiplies the compensation signal CV by -1 and adds the output signal OUT). The sense amplifier 431 is coupled to the current subtractor 432 and compares the subtraction result generated by the current subtractor 432 with a plurality of preset reference currents, and thereby senses the compensated output signal COUT.

In this embodiment, the memory strings 411 and 421 can be set in different positions, but share the same address information. In other words, the memory strings 411 and 421 can be simultaneously written and read at the same time to provide read current at the same time.

Please refer to FIG. 4 and FIG. 5 simultaneously. FIG. 5 is a flowchart of the operation of the embodiment of the present invention in FIG. 4. Step S510 converts multiple weights into multiple transduction values G ₁ , G ₂ and G ₃ , and generates another set of multiple reference transduction values G ^' ₁ , G ^' ₂ and G ^' ₃ according to the transduction values. Wherein, the total of the transduction values G ₁ , G _2, and G ₃ is the same as the total of the reference transduction values G ^′ ₁ , G ^′ ₂ and G ^′ ₃ . And, according to the transduction values G ₁ , G _2, and G ₃ and the reference transduction values G ^′ ₁ , G ^′ ₂ and G ^′ ₃ , write operations are performed on the memory cells MC1 to MC3 and the reference memory cells MCR1 to MCR3, respectively. In addition, step S520 provides the input signals V ₁ to V ₃ to the memory cells MC1 to MC3 to obtain the output signal OUT, and provides the reference input signal V _base to the reference memory cells MCR1 to MCR3 to obtain the compensation signal CV. Step S530 is to perform a sensing action on the difference between the output signal OUT and the compensation signal CV to generate the compensation output signal COUT.

Please refer to FIG. 6. FIG. 6 is a schematic diagram of another implementation of the data sensing device according to the embodiment of the present invention. The data sensing device 600 includes a compensation signal generator 610, a weight operator 620, and an arithmetic operator 630. The compensation signal generator 610 includes a memory string 611 formed by a plurality of reference memory cells MCR1 ˜MCR3. The compensation signal generator 610 performs a write operation on the reference memory cells MCR1 ˜MCR3 according to a plurality of reference weights, and causes the reference memory cells MCR1 ˜MCR3 to provide a plurality of reference conduction values respectively. In addition, the memory string 611 of the compensation signal generator 610 also receives the reference input signal V _base , and provides a read current through the bit line BL1 of the memory string 611 as the compensation signal CV.

On the other hand, the weight calculator 620 has a memory string 621 formed by a plurality of memory cells MC1-MC3. The weight calculator 620 performs write operations on the memory cells MC1~MC3 respectively according to a plurality of weight values, so that the memory cells MC1~MC3 respectively provide a plurality of conduction values. It is worth mentioning that the sum of the reference conduction values is the same as the sum of the conduction values. In addition, the memory cells MC1 to MC3 respectively receive input signals V ₁ , V ₂ , and V ₃ , and provide a read current through the bit line BL2 of the memory string 621 as the output signal OUT.

In this embodiment, the memory strings 611 and 621 may correspond to the same word line, and respectively correspond to different bit lines BL1 and BL2.

In this embodiment, the arithmetic operator 630 includes a current adder 631 and a sense amplifier 632. The current adder 631 receives the compensation signal CV and a plurality of reference currents ref _a , and adds the reference current ref _a and the compensation signal CV to generate a plurality of compensation reference currents ref _b . The current adder 631 also provides the compensation reference current ref _b to the sense amplifier 632. The sense amplifier 632 further receives the output signal OUT, and generates the compensation output signal COUT by _{comparing the output signal OUT with the compensation reference current ref b.}

Please refer to FIG. 7 for a schematic diagram of the offset operation of the reference current according to the embodiment of the present invention. In the embodiment of FIG. 6, the arithmetic operator 630 performs the _{operation by using the reference current ref a of the sense amplifier 632 as the sensing basis.} The offset is used to generate the compensation reference current ref _b , and then by comparing the output signal OUT with the compensation reference current ref _b , the correct compensation output signal COUT can also be effectively generated.

Please refer to FIG. 6 and FIG. 8 simultaneously. FIG. 8 is a flowchart of the operation of the embodiment of the present invention in FIG. 6. Step S810 converts multiple weights into multiple transduction values, and generates another set of multiple reference transduction values according to the transduction values. Among them, the total transduction value is the same as the sum of the reference transduction value. And, according to the transduction value and the reference transduction value, write operations are performed on the memory cells MC1~MC3 and the reference memory cells MCR1~MCR3, respectively. Further, step S820 provides the input signals V ₁ ~ V ₃ to the memory cell MC1 ~ MC3 to obtain an output signal OUT, and provides a reference input to the reference signal V _base MCR1 ~ MCR3 memory cell to obtain a compensation signal CV. In step S830, the compensation signal CV and the reference current ref _a are added to obtain a plurality of compensation reference currents ref _b , and the output signal OUT is sensed according to the compensation reference current ref _b to generate the compensation output signal COUT.

Please refer to FIG. 9 below. FIG. 9 shows a flowchart of a data sensing method according to an embodiment of the present invention. Wherein, step S910 provides a compensation signal generator to receive a reference input signal and a plurality of reference weight values, and generates a compensation signal according to the reference input signal and the reference weight value; step S920 provides a weight calculator, and based on the address information, the weight calculator The multiple memory cells in the memory cell perform write operations according to multiple weight values, and the memory cells receive multiple input signals to generate output signals; and, in step S930, an arithmetic operator is provided to perform operations on the output signal and the compensation signal. Generate output signal after compensation.

The implementation details of the foregoing steps have been described in detail in the foregoing embodiments and implementation manners, and will not be repeated here.

In summary, the present invention provides a compensation signal generator to generate a compensation signal based on a reference input signal and a plurality of reference weight values, and enables the neural operation result (output signal) generated by the weight calculator to generate the compensation signal, and The reference signal performs arithmetic operations to generate the compensated output signal. Since the compensation signal is independent of its input, the compensation signal can be calculated in advance. Therefore, in actual application, it can only take half the amount of calculation compared with the conventional technology. In addition, when multiple weight calculators calculate the same weight, its The compensation signal generator can be shared, so the present invention can use a lower amount of calculation to complete a neural network-like calculation action with a negative input signal.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. The scope of protection of the present invention shall be determined by the scope of the attached patent application.

100, 200, 400, 600: Data sensing device 110, 210, 410, 610: Compensation signal generator 120, 220, 420, 620: Weight calculator 130, 230, 430, 630: Arithmetic calculator 221, 431, 632, SA: Sense amplifier 222, 421, 411, 611: Memory string 432: Current subtractor 631: Current adder BL1, BL2: Bit line V _base : Reference input signal WR: Reference weight value CV: Compensation signal ADD: address information W ₁ ~ W _M : weight value V ₁ ~ V _N : input signal OUT: output signal COUT: output signal after compensation MC1~MCM: memory cell MCR1~MCR3: reference memory cell G ^' ₁ , G ^' _2. G ^' ₃ : Reference conduction value G ₁ , G ₂ , G ₃ : Conduction value ref _a : Reference current ref _b : Compensation reference current S310~S330, S510~S530, S810~S830, S910~S930: Data sensing step

FIG. 1 is a schematic diagram of a data sensing device according to an embodiment of the invention. FIG. 2 is a schematic diagram of an implementation of the data sensing device according to the embodiment of the present invention. Fig. 3 shows a flowchart of the operation of the embodiment of Fig. 2 of the present invention. FIG. 4 is a schematic diagram of another implementation of the data sensing device according to the embodiment of the present invention. FIG. 5 shows a flowchart of the operation of the embodiment in FIG. 4 of the present invention. FIG. 6 is a schematic diagram of another implementation of the data sensing device according to the embodiment of the present invention. FIG. 7 is a schematic diagram of the offset operation of the reference current according to the embodiment of the present invention. FIG. 8 is a flowchart of the operation of the embodiment of FIG. 6 of the present invention. FIG. 9 shows a flowchart of a data sensing method according to an embodiment of the present invention.

100: data sensing device

110: Compensation signal generator

120: weight calculator

130: Arithmetic Operator

V _base : reference input signal

WR: Reference weight value

CV: Compensation signal

ADD: address information

W ₁ ~W _M : weight value

V ₁ ~V _N : Input signal

OUT: output signal

COUT: Output signal after compensation

Claims (12)

A data sensing device, coupled to a memory array, includes: a compensation signal generator that receives a reference input signal and a plurality of reference weight values, and multiplies the sum of the reference input signal and the reference weight values by A compensation signal is generated; a weighting calculator has a plurality of memory cells, the weighting calculator is based on one-bit address information, the memory cells are written according to a plurality of weight values, and the weighting calculator makes the The memory cells receive a plurality of input signals to generate an output signal; and an arithmetic operator, coupled to the compensation signal generator and the weighting operator, generates a compensated signal according to the result of the subtraction of the output signal and the compensation signal The output signal, or the sum of the compensation signal and a plurality of first reference currents, is compared with the output signal to generate the compensated output signal. For the data sensing device described in claim 1, wherein the weighting calculator makes the memory cells receive the input signals and generate a reading current, and the weighting calculator further senses the reading Current to generate the output signal. According to the data sensing device described in claim 2, wherein the weighting operator further includes a sense amplifier to compare the read current with a plurality of second reference currents to generate the output signal. In the data sensing device described in item 1 of the scope of patent application, the memory cells are non-volatile memory cells. According to the data sensing device described in item 1 of the scope of patent application, the memory cells respectively provide a plurality of transduction values according to the corresponding weight values. According to the data sensing device described in claim 1, wherein the compensation signal generator includes a plurality of reference memory cells, the compensation signal generator is based on the address information, and the reference memory cells are based on the reference The weight value is written in, and the compensation signal generator causes the reference memory cells to receive the reference input signal to generate the compensation signal, wherein the sum of the reference weight values is equal to the sum of the weight values. According to the data sensing device described in item 6 of the scope of patent application, the arithmetic operator includes: a current subtractor, coupled to the compensation signal generator and the weight operator, so that the output signal is subtracted from the compensation signal to Generating a subtraction result; and a sense amplifier, coupled to the current subtractor, and generating the compensated output signal based on comparing the subtraction result with a plurality of reference currents. According to the data sensing device described in claim 6, wherein the arithmetic operator includes: a current adder, coupled to the compensation signal generator, so that the compensation signal and a plurality of reference currents are added to generate a plurality of compensations Reference current; and a sense amplifier, coupled to the current adder, based on comparing the output signal with the compensation reference currents to generate the compensated output signal. According to the data sensing device described in item 6 of the scope of patent application, the reference memory cells are non-volatile memory cells. A data sensing method includes: providing a compensation signal generator to receive a reference input signal and a plurality of reference weight values, multiplying the reference input signal and the sum of the reference weight values to generate a compensation signal; providing a The weighting calculator, based on one-bit address information, performs write operations on the plurality of memory cells in the weighting calculator according to a plurality of weight values, and makes the memory cells receive a plurality of input signals to generate an output signal; And providing an arithmetic operator to generate a compensated output signal according to the result of the subtraction of the output signal and the compensation signal, or compare the sum of the compensation signal and a plurality of first reference currents with the output signal to generate the Output signal after compensation. For the data sensing method described in claim 10, the step of making the memory cells receive the input signals to generate the output signal includes: making the memory cells receive the input signals to generate a reading Taking a current; and sensing the read current to generate the output signal. For example, the data sensing method of claim 10, wherein the compensation signal generator is provided to receive the reference input signal and a plurality of weight values, and the sum of the reference input signal and the reference weight values is multiplied by The step of generating the compensation signal includes: setting a plurality of reference memory cells in the compensation signal generator; based on the address information, performing a write operation on the reference memory cells according to the reference weight values; and The reference memory cells are made to receive the reference input signal to generate the compensation signal, wherein the sum of the reference weight values is equal to the sum of the weight values.

Images