TWI794910B - Memory and training method for neutral network based on memory - Google Patents

Abstract

The present invention discloses a memory and a training method for neutral network based on memory. The training method includes: obtaining one or more transfer functions of a memory corresponding to one or more influence factors; determining a training plan according to an ideal case and the one or more influence factors; training the neutral network according to the training plan and the one or more transfer functions to obtain a plurality of weights of the trained neutral network; and programming the memory according to the weights.

Description

Memory and training method for memory-based neural network

The invention relates to a memory and a training method for a neural network based on the memory.

A neural network is a mathematical model that can be used for machine learning. In the past, the neural network is to establish a mathematical model in the form of software on a computer device, and use the processor to perform calculations based on the mathematical model. As the requirements for computing speed increase, the use of memory with high-speed computing capabilities to implement neural networks has become the focus of research. This approach is called computing in memory. Although in-memory computing has many advantages, due to the characteristics of memory, such as being affected by temperature, read and write times, etc., the relationship between the input and output of memory cells is usually not ideally linear. How to reduce the adverse effect of non-ideal memory on the correctness of decision-making based on the results calculated in memory is an important issue.

An embodiment of the present invention discloses a training method for a memory-based neural network. The training method includes: obtaining one or more conversion functions corresponding to one or more influencing factors of a memory; determining a training plan according to an ideal situation and the one or more influencing factors; according to the training plan and the One or more transformation functions for training the type of neural network to obtain a plurality of weights of the type of trained neural network; and programming the memory according to the weights.

Another embodiment of the present invention discloses a memory. The memory includes a plurality of memory cells to represent a plurality of synapses of a type of neural network. Each memory cell includes a resistor. The resistance values of the resistors are determined according to the following methods: obtain one or more transfer functions corresponding to one or more influencing factors of the memory cells; determine a training according to an ideal situation and the one or more influencing factors program; according to the training program and the one or more transfer functions, the type of neural network is trained to obtain a plurality of weights of the synapses of the trained type of neural network; and according to the The weight determines the resistance values of the resistors of the memory cells.

In order to have a better understanding of the above-mentioned and other aspects of the present invention, the following specific examples are given in detail with the accompanying drawings as follows:

Please refer to Figure 1, which shows a schematic diagram of an operational memory. The operation memory 10 includes a plurality of word lines WL1-WL3, a plurality of bit lines BL1-BL3, and a plurality of memory cells C11-C33. Each memory cell C11-C33 is coupled to a corresponding word line and a corresponding bit line. The memory cells C11-C33 may respectively include a variable resistor, and the resistance values are R11-R33 respectively. In one embodiment, the variable resistor can be realized by a transistor. The operation memory 10 can realize a type of multiplication and addition (sum of product) of a neural network through operations in the memory. The operation principle of the computing memory 10 is described in detail as follows.

Memory cells C11~C33 can be used to represent synapses in a neural network. The resistance values R11-R33 or conductance values (ie the reciprocal of the resistance value) of the variable resistors of the memory cells C11-C33 can be used to represent the weight of the synapse. The input voltages V1-V3 of the word lines WL1-WL3 can be used to represent input data. The output currents I1~I3 of the bit lines BL1~BL3 are respectively: I1=V1*1/R11+V2*1/R21+V3*1/R31, I2=V1*1/R12+V2*1/R22+V3 *1/R32, I3=V1*1/R13+V2*1/R23+V3*1/R33. The currents I1~I3 represent the results of the input data after the neural network-like operation. The currents I1 - I3 can be sent to a decision circuit (not shown) for subsequent operations to generate a decision.

It should be noted that the example in FIG. 1 is simplified for convenience of description. In practical applications, the number of memory cells of the computing memory is usually very large, and the architecture of the computing memory will also vary with the neural network to be implemented.

In order to improve the accuracy of decision-making, it is usually necessary to train the quasi-neural network. For example, input the training data into the neural network for calculation, generate a decision based on the calculation result of the neural network, compare the decision with the answer (ground truth) corresponding to the training data, and compare the neural network based on the comparison result. Adjust the weights in . There are currently two training methods for memory-based neural networks: (1) using computer devices and software to train according to the ideal mathematical model of neural networks, and (2) actually training on memory. However, method (1) does not take into account the non-ideal characteristics of the memory, and method (2) does not take into account the ideal mathematical model. The neural network trained by the two methods runs on the memory, which will make the subsequent decision-making accuracy insufficient. In order to solve the above problems, the present invention proposes a training method for memory-based neural network.

Please refer to FIG. 2 , which is a flow chart of a training method according to an embodiment of the present invention. This training method can be applied to memory-based neural networks.

In step S201, one or more transfer functions corresponding to one or more influencing factors of a memory corresponding to a type of neural network are acquired. The memory corresponding to the neural network means that the memory is used as hardware to realize the neural network. The conversion function represents the relationship between the input and output of the memory cell of the memory corresponding to the impact factor. Influencing factors may include different temperatures, different read and write times, different input voltages, changes in each read operation, memory retention capability, and write error, etc. The so-called memory retention capacity refers to the time that information is stored in memory cells. The so-called written value error refers to the difference between the actual value written during programming and the ideal value. For example, the impact factor may include a temperature of 40 degrees Celsius, a temperature of 50 degrees Celsius, 5,000 times of reading and writing, and 10,000 times of reading and writing. For actual memory, elements such as transistors and variable resistors included in memory cells are usually not ideal. This imperfection will cause the ratio of the input voltage to the output current of the memory cell to be affected by the influencing factors rather than a fixed value.

Please refer to FIG. 3, which shows the relationship between the input voltage and the output current of a memory cell corresponding to a specific temperature. In Fig. 3, the horizontal axis is the input voltage of the memory cell, the vertical axis is the output current of the memory cell, and the multiple curves are different transfer functions, corresponding to multiple resistance values to be programmed for the memory cell. In an ideal situation, a specific weight corresponds to a specific resistance value, and the resistance value should not change without changing the weight. That is to say, under ideal conditions, the ratio between the input voltage and the output current should be a constant value, that is, the transfer function should be a straight line. However, under actual conditions (non-ideal conditions), the ratio of input voltage to output current will vary with the input voltage. Therefore, it can be seen from Fig. 3 that the transfer function under actual conditions is not a straight line.

It should be noted that the conversion function can be obtained through actual testing, according to historical records, or in any other way. The present invention is not limited.

In step S203, a training plan is determined according to the one or more influencing factors and an ideal situation. The ideal situation represents that the memory is not affected by any influencing factors and fully conforms to the ideal mathematical model of the neural network. In other words, ideally, for each memory cell, the ratio of the input voltage to the output current corresponding to a specific resistance value is constant. In one embodiment, determining the training plan may include steps S401-S407, as shown in FIG. 4 . In step S401, an ideal condition or an unideal condition is determined. Step S403 is executed when it is determined to be an ideal situation, and Step S405 is executed when it is determined to be an unideal situation. In step S403, the training times are determined, and the training plan is updated. In step S405, the influencing factor is determined. That is, step S403 is executed after one of the influencing factors is selected to determine the training times for the training factor. In step S407, it is determined whether the configuration of the training plan is completed, if the determination is no, return to step S401, if the determination is yes, the process ends. For easier understanding, a practical example is given below as an illustration. First execute steps S401 and S403 to add the ideal situation and the number of times to the training plan as five times, then return to step S401 to determine the non-ideal situation, and execute steps S405 and S403 to determine the influencing factor as the temperature of 40 degrees Celsius and the number of times as ten Then return to step S401 to determine the unsatisfactory situation and execute steps S405 and S403 to determine the impact factor is 10,000 reading and writing times and 7 times and add to the training plan, then end this book process. In this way, the training plan can be obtained as: [(ideal situation, five times); (temperature 40 degrees Celsius, ten times); (reading and writing times 10,000 times, seven times)]. In one embodiment, the content of the training plan can be further scheduled. For example, the training plan after scheduling in the aforementioned example can be: [(ideal condition, twice); (temperature 40 degrees Celsius, five times); (ideal condition, once); (reading and writing times 10,000 times, Three times); (ideal situation, one time); (10,000 reading and writing times, four times); (ideal situation, one time); (temperature 40 degrees Celsius, five times)]. That is to say, the training for ideal conditions and different influencing factors can be interleaved, and the training for the same influencing factor does not need to be completed consecutively. Interspersed training for different impact factors can prevent the neural network from being too affected by specific impact factors, so that multiple impact factors can contribute more evenly to the neural network. The actual training plan can adaptively configure the occurrence frequency and training times of each influencing factor according to the actual characteristics of the memory. For example, when the memory is more sensitive to temperature, by increasing the frequency and training times of the temperature-influencing factor in the training plan, the trained neural network can be better able to cope with the impact of temperature changes.

In step S205 , according to the training plan and the conversion function associated with the training plan, the neural network is trained with a computer device and software simulation, so as to obtain a plurality of weights of the trained neural network. Simulation refers to the execution of neural network-like mathematical model operations by the processor of a computer device, and the simulation of non-ideal conditions with software assistance. The so-called software-assisted simulation of non-ideal conditions means that each time the training corresponding to the impact factor is performed, for each synapse in the neural network, according to the conversion function corresponding to the impact factor, the weight corresponding to the synapse, and the input The input value of the synapse adjusts the output value of the synapse. In one embodiment, the training of the neural network can be performed sequentially according to the training plan. In another embodiment, the training of the neural network can be performed in any order according to the training plan. Generally speaking, a single training process includes: input training data into the neural network; obtain one or more decision parameters by calculating according to the training data through the neural network; generate a decision according to the decision parameters; calculating a score on a solution of the training data; and adjusting one or more weights of the neural network based on the score. Since there are many existing technologies for the details of the training process, no further description is given here.

In step S207, the memory cells of the memory are programmed according to the weights to obtain a memory (operational memory) for realizing a neural network.

It should be noted that any suitable memory type can be applied to the training method of the present invention. For example, phase-change memory (phase-change memory, PCM), ferroelectric random access memory (Ferroelectronic RAM, FeRAM), etc. can all be applicable to the training method proposed by the present invention.

In addition, according to the above training method, the present invention also proposes a memory. A memory has multiple memory cells. Each memory cell includes a resistor. These memory cells are used to represent multiple synapses of a type of neural network. The resistance values of the resistors of the memory cells are determined according to a plurality of weights obtained through the above training method.

Please refer to Table 1. In Table 1, ACCi is the correctness of the decision obtained by simulating the trained neural network under ideal conditions according to the computer device, and ACCn is the correctness of the decision based on the memory used to realize the trained neural network. the correctness of the resulting decision. It can be seen from Table 1 that in the five tests, the gap between the correctness of the ideal situation and the correctness of the internal memory operation is within 2%. This means that the training method proposed by the present invention can effectively improve the problem that the correctness of subsequent decision-making is reduced due to the characteristics of the memory itself when the memory is used to implement the neural network. ACCi ACCn |ACCi-ACCn|/ACCi test 1 0.8756 0.8818 0.71% test 2 0.8668 0.8735 0.77% test 3 0.8683 0.8745 1.55% test 4 0.8731 0.8651 0.78% test 5 0.8585 0.8529 0.65% Table I

To sum up, although the present invention has been disclosed by the above embodiments, it is not intended to limit the present invention. Those skilled in the art of the present invention can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be defined by the scope of the appended patent application.

10: Operational memory C11~C33: memory cells WL1~WL3: character line BL1~BL3: Bit lines S201~S207, S401~S407: steps

FIG. 1 shows a schematic diagram of an operational memory. FIG. 2 shows a flowchart of a training method according to an embodiment of the present invention. Fig. 3 shows a schematic diagram of the transfer function. FIG. 4 shows a flow chart of determining a training plan according to an embodiment of the present invention.

S201~S207: steps

Claims (10)

A training method for a memory-based neural network comprising: obtaining one or more transfer functions corresponding to one or more influencing factors of a memory; determining a training program according to an ideal situation and the one or more influencing factors; training the neural network according to the training plan and the one or more transformation functions to obtain a plurality of weights of the trained neural network; and The memory is programmed according to the weights. The training method as claimed in claim 1, wherein the one or more influencing factors include temperature, reading and writing times, input voltage, variation of each reading operation, memory retention capacity and writing error. The training method as described in claim 1, wherein determining a training plan according to an ideal situation and the one or more influencing factors includes: Determine a first training times for the ideal situation, and update the training plan; Select at least one of the one or more influencing factors, determine a second training times corresponding to each of the selected one or more influencing factors, and update the training plan. The training method as described in Claim 1, wherein the neural network of this type is trained according to the training plan and the one or more transformation functions, so as to obtain the plurality of weights of the neural network of the trained type, The training for the influencing factor of the decision is performed according to the one or more transfer functions corresponding to the influencing factor of the decision, the plurality of inputs of the plurality of synapses of the neural network of the type, the plurality of synapses of the synapses The weights to be trained adjust the plural outputs of these synapses. The training method according to claim 3, wherein in the training plan, the training corresponding to the ideal situation is interspersed with the training corresponding to the one or more influencing factors. A memory comprising: A plurality of memory cells are used to represent a plurality of synapses of a type of neural network, each of the memory cells includes a resistor, and the resistance value of the resistors is determined according to the following method: Obtain one or more transfer functions corresponding to one or more influencing factors of the memory cells; determining a training program according to an ideal situation and the one or more influencing factors; training the neural network according to the training plan and the one or more transfer functions to obtain a plurality of weights of the synapses of the trained neural network; and The resistance values of the resistors of the memory cells are determined according to the weights. The memory according to claim 6, wherein the one or more influencing factors include temperature, read/write times, input voltage, variation per read, memory retention capability, and write error. The memory as described in claim 6, wherein determining a training plan according to an ideal situation and the one or more influencing factors includes: Determine a first training times for the ideal situation, and update the training plan; Select at least one of the one or more influencing factors, determine a second training times corresponding to each of the selected one or more influencing factors, and update the training plan. The memory as described in claim 6, wherein the neural network of this type is trained according to the training plan and the one or more transformation functions, so that among the plurality of weights of the neural network after training, The training for the influencing factor of the decision is performed according to the one or more transfer functions corresponding to the influencing factor of the decision, the plurality of inputs of the plurality of synapses of the neural network of the type, the plurality of synapses of the synapses The weights to be trained adjust the plural outputs of these synapses. The memory according to claim 8, wherein in the training plan, the training corresponding to the ideal situation is interspersed with the training corresponding to the one or more influencing factors.

Images