KR20210017835A - Neuromorphic Memory Management System and Memory Allocation Method thereof - Google Patents

Abstract

According to one aspect of the present invention, a memory management system comprises: a memory having a multidimensional lattice structure; a prediction unit predicting the memory usage level of each of two or more data sets and determining the order to be inputted to the memory; and a management unit allocating a memory area to store the two or more data sets according to a prediction result of the prediction unit to store, learn, and infer the data sets. The memory has one of the two-dimensional and three-dimensional lattice structures, and the memory can be accessed from two or more directions according to each dimension. Therefore, the power consumption can be reduced.

Description

Neuromorphic Memory Management System and Memory Allocation Method thereof

The present invention relates to memory management, and more particularly, to a system and method for managing a memory in a neuromorphic semiconductor within a low power device or a device requiring independent operation.

The synapse of a neuromorphic device is composed of an array structure, and a learning function is implemented and data is calculated using various memory units.

Neuromorphic memory for this purpose generally stores 128 bytes of data per neuron, and the conventional memory allocation method sequentially stores and processes data from the first location of the memory in a daisy chain method as shown in FIG. 1. For example, 640 bytes of A input data uses 5 neurons 0-4 for learning and inference.

Since conventional neuromorphic memory receives and processes one input data at a time, if B input data (640 bytes) is to be processed subsequently, the entire memory (neuron) is reset first, and then B input data is converted to A input data. Input to neurons 0-4 used by

In the end, in the conventional method, when the application or service data to be processed is heterogeneous or different data, to learn data B (indicated in green) as shown in FIG. 2, the previously stored and processed application data A (indicated in yellow). ) Must be reset. At this time, the learning contents of application data A are all stored in the external memory unit for later recognition (inference), and due to this process, new data is updated according to the type, processing speed, and access speed of the external memory unit. There is a lot of difference in data processing performance of the Lomorphic memory.

In order to obtain good processing speed, the speed of the system bus and fast data exchange between the external memory unit and the system core are required. Therefore, the above data rate is significantly lowered for low-power/low-spec devices, which may cause a data bottleneck. There may be limitations.

Moreover, the standard of power consumption or performance may be dependent on an external memory unit rather than an actual neuromorphic semiconductor.

This problem occurs not only in the learning stage, but also in the process of retrieving the learned result from an external memory for recognition or inference of input data.

On the other hand, in terms of utilizing the memory in the actual neuromorphic semiconductor, the conventional neuromorphic memory also has the characteristic of sequentially storing and processing data from the first location of the memory in a daisy chain method as shown in FIG. Therefore, depending on the application, when there are many unused memories, inefficiency may occur in terms of memory usage.

More specifically, as shown in FIG. 1, after data input is completed, comparison parallel operation can be performed simultaneously as indicated by a blue arrow, but the method of writing data to memory for parallel operation in a neuromorphic semiconductor Together, they are using the daisy chain method. Therefore, when there is little data to be stored for learning or inference, the remaining neuron memory is not used, and as shown in the example of FIG. 1, only a very small area of memory is used and most of the remaining areas may not be used. .

In this way, in terms of utilizing the memory in the neuromorphic semiconductor, there is a problem in that it is difficult to recycle the unused memory portion.

On the other hand, due to the characteristics of the conventional neuromorphic semiconductor, there is a problem that the method for recursive operation is not applied, so the algorithm used for deep learning or the algorithm requiring regression operation in machine learning cannot be applied. There is.

An object of the present invention is to provide a memory management system and method capable of variably changing a memory structure in a neuromorphic semiconductor, deviating from the conventional neuromorphic memory allocation method of a uniform daisy chain method.

Another object of the present invention is to provide a memory management system and method for increasing memory use efficiency by allocating memory based on predictions about memory usage and allocating to use a memory area that is not currently used.

In particular, when the expected memory size is incorrect, we intend to apply a space utilization function that can be readjusted through feedback.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problem as described above, and other technical problems may exist.

In the present invention, by changing the memory structure in the neuromorphic semiconductor, the Prescience Estimator (PSE) manager predicts the degree of memory usage for the application, and allocates a part from the entire neuromorphic memory structure of the Prescience Manager (PSM) to solve the above problems. Achieve.

A memory management system according to an aspect of the present invention includes a memory having a multidimensional lattice structure, a prediction unit that predicts a memory usage degree of each of two or more data sets and determines an order of input to the memory, and a prediction result of the prediction unit. And a management unit for storing, learning, and reasoning by allocating a memory area to store the at least two data sets according to the data set.

The memory has one of the two-dimensional and three-dimensional lattice structures, and is a neuromorphic memory that can be accessed from two or more directions according to each dimension.

The prediction unit predicts the type and size of data input in a predetermined time interval based on one or more of pre-stored reference information and application output data information to be executed, and for the two or more data sets, the predicted size order of each data set The order of memory allocation is determined.

In addition, the prediction unit may determine the order of memory allocation and determine the memory area based on the predicted size of each data set and the number of cells on one side of the memory area temporarily set for each data set. Perform reallocation.

A memory management system according to another aspect of the present invention includes a memory having a multidimensional lattice structure, a prediction unit that predicts a memory usage degree of each of two or more data sets, and the two or more data sets according to a prediction result of the prediction unit. And a management unit for storing, learning, and inferring the data set by allocating a memory area to be stored according to a lambda element technique.

The memory includes a storage unit that stores the two or more data sets and reference data to be used for calculation, and an operation unit that calculates the data set based on the reference data.

The management unit performs learning on the two or more data stored in the memory and controls the memory to store the learning result, and alternately stores the two or more data in two or more axial directions of the memory.

A method for managing a neuromorphic semiconductor memory according to another aspect of the present invention includes the steps of allocating and storing memory regions for two or more input data sets, performing learning on the two or more input data sets, and And storing reference data as a learning result in a new memory area, and performing a comparison operation with the reference data on one or more new input data sets.

Storing the input data includes predicting the size of each of the two or more input data sets, and first allocating and storing a memory area for an input data set having a larger size among the input data sets according to the prediction result. Is performed by

Meanwhile, the step of determining reference data to be used for a subsequent calculation of another new input data set among the stored reference data may be further included according to the result of the calculation step.

The prediction-based memory management system according to the present invention can reuse the reference data by re-allocating the comparison decision value of the single layer method to another memory at a time difference.

In addition, since it is performed in a core system inside a neuromorphic semiconductor without accessing an external memory, processing speed is fast and power consumption is reduced accordingly.

According to the present invention, since a multi-layer method and a weight method capable of regression operation can be simultaneously used using a neuromorphic memory, various deep learning algorithms can be optimized from a simple algorithm application method through a single layer due to the characteristics of the existing neuromorphic semiconductor. It shows the advantage of being able to apply the algorithm of a machine learning algorithm (e.g., Support Vector Machine (SVM)).

In addition, there is an advantage in that an unused area can be recycled by allocating a neuromorphic memory that does not use input data, and various data can be learned at once without a reset process in a single input. Therefore, since various data are learned at a time, one recognition rule can be defined using fused information, and since one storage of the recognition rule is required, there is an advantage of reducing the number of storage of external memory units by the system bus.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a diagram showing a structure of a neuromorphic memory and a method of storing input data in a conventional neuromorphic memory.
2 is a diagram showing a method of storing a plurality of different input data in a conventional neuromorphic memory.
3 is a block diagram of a memory management system according to the present invention.
4 is a diagram showing a three-dimensional implementation example of a variable memory according to the present invention.
5 is a diagram illustrating an aspect in which two or more data of a variable memory according to the present invention are simultaneously stored together.
6 is a diagram showing an aspect of allocation of two or more data in a variable memory according to the present invention.
7 is a diagram showing a memory management system according to the present invention.
8 is a diagram showing a basic method of a memory allocation method according to the present invention.
9 is a diagram illustrating an optimal memory allocation pattern according to the present invention.
10 is a diagram showing an example of storing two or more types of data in one memory at the same time to perform a data operation according to the present invention.
11 is a conceptual diagram for a multiple data operation method according to the present invention.
12 is a conceptual diagram of a method of calculating input data over time according to the present invention.
13 is a conceptual diagram for a complex operation method according to the present invention.
14 is a view for explaining an operation method for increasing precision step by step according to the present invention.

Advantages and features of the present invention, and a method of achieving them will become apparent with reference to the embodiments described in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various forms different from each other, only this embodiment makes the disclosure of the present invention complete, and those skilled in the art to which the present invention pertains. It is provided to fully inform the scope of the invention to the person, and the invention is defined by the description of the claims.

On the other hand, terms used in the present specification are for explaining examples and are not intended to limit the present invention. In this specification, the singular form also includes the plural form unless specifically stated in the phrase. As used herein, "comprises" or "comprising" refers to the presence of one or more other components, steps, actions and/or elements other than the recited elements, steps, actions and/or elements, or Does not rule out addition.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

Meanwhile, when a certain embodiment can be implemented differently, functions or operations specified in a specific block may be executed differently from the order specified in the flowchart. For example, two consecutive blocks may actually be executed at the same time, or the blocks may be executed in reverse depending on a related function or operation.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In adding reference numerals to elements of each drawing, even though they are indicated on different drawings, the same elements are assigned the same reference numerals as much as possible, and in describing the present invention, a detailed description of related known configurations or functions If the gist of the present invention may be obscured, a detailed description thereof will be omitted.

The basic technical idea of the present invention is to enable a memory write structure in the form of a memory grid by using a variable memory structure in order to use the above-described memory reallocation structure.

To this end, the Prescience Estimator (hereinafter referred to as PSE in this specification) predicts the degree of memory usage for the application, and based on this, the Prescience Manager (hereinafter referred to as PSM) A memory management system that allocates a part of the entire neuromorphic memory structure is provided to reuse a memory area that is not used by other applications to increase the efficiency of memory use.

A memory management system according to an exemplary embodiment of the present invention will be described with reference to FIG. 3.

The neuromorphic memory management system 300 includes a PSE 320 (Prescience Estimator) that predicts the degree of memory use for each application that will use the memory 330, and a PSM that allocates part or all of the entire neuromorphic memory based on this. (310; Prescience Manager). PSE and PSM can be implemented as software modules inside the system core of the neuromorphic chip, but of course they can also be implemented as separate hardware.

The memory 330 has a multidimensional lattice structure, and has a variable memory structure in which memory access is possible in various directions according to each dimension. For example, if you have a two-dimensional grid structure, you can access memory in four directions, up, down, left, and right, and if you have a three-dimensional grid structure, you can access them in both directions (a total of six directions) on each axis of x, y, and z. do.

The variable memory structure may be implemented by configuring an existing neuromorphic memory to have a logical multidimensional structure, and constructing a three-dimensional structure by stacking two or more neuromorphic memories in a stack structure, or as shown in FIG. Likewise, it can be implemented in a three-dimensional structure according to a semiconductor process method.

As shown in FIG. 5, memory allocation is configured to have a two-dimensional plane or three-dimensional dimensional structure (single lattice or multiple lattice structure), unlike a one-dimensional structure (daisy chain method) of a conventional neuromorphic semiconductor. As shown in FIG. 5, the variable memory according to the present invention has a structure in which data is allocated along a blue line and memory is allocated along an orange line, which is another path.

As a result. As shown in FIG. 6, several types of data may be simultaneously stored in the neuromorphic memory, and several types of data may be processed at the same time.

Fig. 7 schematically shows the effective utilization and effect of such a variable memory structure. As shown, the variable multidimensional memory 300 can be written and read in the horizontal, vertical, and height directions under the control/management of the PSM 310 and the PSE 320.

The advantages that can be obtained by adopting such a memory structure are as follows.

In using the memory space, it is possible to predict the degree of memory use for the application, allocate only a portion suitable for the degree of use in the entire neuromorphic memory structure, and reallocate the unused area for use by other applications. have.

Due to the variable allocation memory of the lattice structure, data storage for learning and comparison operations can be performed at the same time, thereby preventing performance degradation due to access to an external memory unit.

In the case of storing and learning input data with a time difference in a lattice-type variable memory structure, a weighting method capable of regression operation can be applied, so that the effect of various deep learning algorithms can be seen rather than a simple algorithm application through a single layer. Algorithm can be applied. Algorithms to which multiple layers are applied can be simulated and used according to the number of logically or physically stacked memory grids, or by multiplexing data allocation spaces simultaneously processed even within a single grid.

In this way, each data is learned according to the memory area secured through the PSM 310 and PSE 320, and since information fusion learning about various data occurs at once, it is learned even with a small memory size in the neuromorphic chip. This is possible, and fast memory access is possible with low power. Learning results are stored separately and used for future learning recognition.

In the recognition (inference) step, since the PSM 310 knows the state of the data in the learned area, the PSM 310 performs all the necessary procedures for performing the comparison operation of the corresponding data so that when various types of input data are input for recognition. Proceed.

A memory allocation method according to the present invention performed by the PSM 310 and the PSE 320 will be described in detail with reference to several examples below.

Basically, the size of the input data set is allocated in the order of the largest.

First, the PSE 320 determines the usable memory space. The usable memory space may be determined immediately before the time when data input is required, but it is desirable to perform it at all times for stability of the load.

Then, the size of input data sets is predicted. For example, it is predicted how many pieces of data that need to be fused for one determination, for example, detection data for four categories of temperature, humidity, dust, and CO2 for fire detection, will be input in a predetermined time interval.

The prediction method is a table of data types and sizes for each application accumulated in advance according to the type of application currently being executed (e.g., in the case of an indoor fire monitoring application, 5 types of gas concentration data are sampled once per second, or each It can also be estimated based on basic information such as how many times per second of each type of gas concentration data is sampled), or it can be estimated by receiving the type and size of input data for each application being executed.

Alternatively, based on the type and size of data inputted in advance for a predetermined time, the type and size of data to be input thereafter may be (equally) estimated. The prediction method is not limited to the above examples, and the technical It covers the various ways possible within the scope of thought.

When the PSE 320 predicts the size of each input data set, the PSM 310 allocates a memory area to store the input data set according to the predicted size of the input data set. It is preferable to allocate and store memory areas in order of size of each input data set.

Further, the memory management system 300 may verify whether each input data set is properly allocated to the memory and perform reallocation.

To this end, if the PSM 310 sets each memory area in the neuromorphic memory according to the size of the memory size B required for each input data, the PSE 320 determines the memory size B required for each input data. Depending on the size, the utilization threshold (UT = B/) is compared with the number of cells (H) corresponding to the available columns or rows (one side of the memory area temporarily set for each input data) for each input data in the neuromorphic memory. H) is determined, and then the H/UT value is calculated.

If H/UT is 2, it can be considered to be stored in a memory area with a horizontal and vertical area of 1:2, and if it is 1, it can be regarded as being stored in a square area. That is, it is determined that the closer to 1, the greater the degree of clustering in nearby memory cells.

On the other hand, if the UT is less than 1, it can be estimated that there may be wasted memory cells because input data is less than the number of memory cells corresponding to one row or column.

Therefore, it is preferable to set H so that the ratio of H/UT for each input data is close to 1, and it is preferable not to exceed 10, and it is more preferable to be in the range of 1 to 5.

The memory management system 300 according to the present invention reallocates the memory based on the above H/UT value, and calculates the H/UT again according to the memory area and H value for each data. A method of allocating a memory area by repeating the above-described process to a level close to 1 to 5 to finally set the memory area, or repeating the H/UT calculation process in the reallocation process so that each memory allocation area is within 10 Can take.

When a memory area to store each data set is determined, data is stored in each area in the order of the size of the UT or the size of the input data set.

In this way, optimal memory utilization efficiency can be obtained.

Another embodiment of memory allocation according to the present invention will be described with reference to FIGS. 8 and 9.

Four sides of the memory are first allocated to the four input data sets starting from the largest input data set in the order of the size of the input data set, and then the allocation area is set using a gamma element technique. In the case of a 3D memory, 6 sides of the memory are first allocated from the largest input data set to the 6 input data sets, and the remaining allocated areas are set using a gamma element technique. The following description takes the case of a two-dimensional lattice memory as an example, but it can be applied to a high-dimensional memory of three or more dimensions.

The specific method is as follows.

), 즉 메모리 영역의 각 변의 메모리 셀의 수를 확인한다. 1) Check the size of a single memory neuron size and number of neurons along the x-axis and y-axis (

), that is, check the number of memory cells on each side of the memory area.

만족하는 것이 좋다. 즉, 각 메모리 영역은 정사각형에 근사한 영역이 할당되도록 하는 것이 좋다. At this time, if possible

It is good to be satisfied. That is, it is preferable to allocate an area approximate to a square for each memory area.

2) As shown in Fig. 8, the edges of the four sides of the memory are filled in the order in which input data A, B, C, D can be input as much as possible.

엘레먼트 기법을 사용하여 ,

,

수식에 따라 각각 할당하고, ,

,

와 같은 방식으로, 순차적으로 해당 면적을 설정한다.3) After that, there is no overlapping chip for the remaining part of the center and to use it to the maximum

Using the element technique,

,

Assign each according to the formula, and,

,

In the same way as, set the area sequentially.

4) Allocate data in the order of size of each data to the area specified above.

4-1) The first four data A, B, C, and D filling the edges of the four sides are allocated in the order of their size.

,

,

,

의 영역을 채운다. 4-2) After that, in the order of input data size

,

,

,

Fill the area.

>

>

>

의 순서대로 해당 영역에 저장한다.FIG. 9 is a diagram illustrating an optimal data input method according to a third embodiment of the present invention. In the case of FIG. 9, first, the edges of the four sides are in this order according to the size of A>B>C> D. Fill, then for input data with the next size

>

>

>

Save in the corresponding area in the order of.

To this end, the PSM 310 and the PSE 320 must all know or predict (estimate) the size of each input data.

In this embodiment, if the size of a data set becomes larger than the memory area allocated in the corresponding order during memory allocation, the data set is first divided and stored in the allocated memory area, and the size of the remaining portion that is not stored It stores the rest of the data set in the smallest of the other, larger memory areas. If the size of the data set is larger than the sum of the sizes of the two memory areas, the data is stored in three or more memory areas in the same manner as above.

When storing one data set in one or more memory areas, the PSM 310 always identifies and manages in which memory areas the data set is stored in one or more memory areas.

In common in the above-described embodiments, memory is allocated and stored in the order of size of input data, and the memory area of each input data is set as close as possible to a square, and the area is set and allocated in the direction of clustering of the same type of input data. Take the way to do it. In other words, the memory area is set so that the remaining unallocated area becomes the largest cluster area possible among the number of cases.

In other words, it is intended to allocate memory to an area having the largest usable area from input data having a large size, thereby minimizing unused memory space.

In addition, the PSE 320 provides a software area to define rules required by the user. That is, an area in which a user can define which combination satisfies which condition for a plurality of inputs is prepared, and when the condition is completed, the PSM 310 determines how and in which memory the corresponding data is to be allocated.

The PSM 310 stores input data in a memory based on the prediction result of the PSE 320, and secures an optimal memory usage area by alternately storing input data for the horizontal and vertical arrays of the 2D memory. . In the case of a 3D memory, input data is alternately stored along an array of horizontal, vertical, and height (ie, three axes of x, y, and z).

In this way, each data is learned according to the memory area secured through the PSM 310 and PSE 320, and information fusion learning for various data occurs at once, so that the memory size of the neuromorphic semiconductor chip It has the advantage of being able to learn even in a small memory, and quick access with low power is possible.

Learning results are stored separately and used for future learning recognition.

In the recognition (inference) step, since the PSM 310 knows the state of the data in the learned area, the PSM 310 performs all the necessary procedures for performing the comparison operation of the corresponding data so that when various types of input data are input for recognition. Proceed.

As described above, in using the memory space, the memory management system 300 according to the present invention estimates the degree of memory use for the application by using the PSM 310 and PSE 320, and in the entire neuromorphic memory structure. Increases memory efficiency by allocating or reallocating unused areas to maximize use.

Compared to the conventional method of using an external memory unit, since data storage for learning and comparison operation for inference can be performed on several data sets at the same time, there is an advantage in that performance degradation due to access to the external memory unit can be prevented.

Hereinafter, a data operation method that can be performed by the memory management system 300 according to the present invention will be described in detail.

Conventional neuromorphic memory stores data in a daisy-chain method as described above, and can then operate on only one type of data at a time. As a result, a single layer is used, and a simple one-time data between input data and training data. Only comparison is possible. As a result, there is a disadvantage that the accuracy of the comparison result (inference result) is very poor.

Of course, even in the single layer method of learning and recognition, the recognition rate tends to increase when the amount of comparison is large. However, due to the nature of the memory manufacturing process, when the amount of memory is increased, manufacturing cost increases and it is difficult to maintain low power. In particular, a problem arises in that sufficient input data required for comparative learning must be secured and pre-processed.

In addition, when learning is necessary, not one input data, but a plurality of input data (for example, when you want to learn temperature data after axis sensor data), in the conventional method, because of the single layer method, one data is learned. There is a problem that a separate process of storing the learned result in an external memory, learning new data again, and storing and comparing the learned result in an external memory is required.

In addition, due to the characteristics of a single layer parallel comparison operation, there is a limitation in that various algorithms required for machine learning cannot be applied. For example, an algorithm such as a support vector machine (SVM) must have a structure in which a plurality of layers is applied and weight calculation is possible.

Accordingly, in the data operation method of the present invention, first, two or more data are stored in different memory areas as shown in FIG. 10 using the above-described variable memory structure and memory management system 300.

In the case of using such a variable structure, as shown in FIG. 11, when the input data is A, B, or A, B, C or more, the effect of a regression operation may be obtained.

This method can be applied to a plurality of different types of data, but it is also possible to track and compare changes in single data over time. 12 shows a case of tracking time difference change of single data, and since weight calculation is possible, not only accuracy improvement but also a regression calculation algorithm can be applied.

In addition, the embodiment shown in FIG. 11 and the embodiment shown in FIG. 12 can be modified in a combined form, and it is also possible to track and compare changes according to time differences of a plurality of data. 13 illustrates a case of tracking change in time difference of such a plurality of data, and various comparison algorithms may be applied.

Hereinafter, some specific embodiments of the calculation method will be described in more detail, but it is a matter of course that the calculation method according to the present invention is not limited thereto.

[Embodiment 1 of the calculation method]

When a plurality of different data are simultaneously input as shown in FIG. 11, for example, when implementing a service that prevents the driver from drowsy driving, the heart rate and the blinking of the driver's biometric information are input as A and B data, respectively, and learn. Then, in the inference step, the comparison operation is performed, and the result is compared with the driving time data (C data), and the comparison operation is performed again to improve accuracy. That is, not uniformly determining whether a person is drowsy only with a certain number of heartbeats and/or blinking, but determining whether he is drowsy by adding a driving time condition to the comparison result of the heartbeat and/or blinking.

For example, the heart rate and blinking when the driving time is less than 1 hour may have different characteristics from the heart rate and blinking when the driving time is 3 hours or more. For example, when driving for a long time, the heart rate and blinking may be lowered or vice versa compared to the state in which the driving started shortly.

Therefore, after learning in advance a plurality of heart rates and number of blinks in the range that can be seen as a drowsy state (data A, B), and learning the heart rate and blink rate that can be seen as a drowsy state by driving time, Primaryly, if the heart rate and blinking are not within the drowsiness range, the comparison operation is not performed on the C data, but if the heart rate and blinking are within the drowsiness range, the comparison operation is performed on the C data and drowsiness by driving time. It determines whether the heart rate and blinking of eyes can be seen as a result of the determination of drowsiness.

Or, if there is a correlation or preferential determination for A and B data, only one of A and B is first compared, and if it is determined that it is in the drowsy area, then other data are compared, and the comparison result of the other data is also In the case of the drowsiness region, the C data may be compared, and each data may be sequentially compared and calculated to be determined.

It would be desirable to learn these A, B, and C data for each driver, but if individual learning is difficult or inconvenient, after storing the reference data (learning value) of the data of several drivers, the range of the corresponding input data is stored. It can be configured to determine whether an area or each condition is satisfied.

In addition, as shown in FIG. 12, an operation for recognizing a situation may be performed according to a change in input data according to a time difference. For example, in the case of data that requires a certain pattern over time to determine a situation, it can be stored by time and applied to check whether the data changes to fit a certain pattern, and according to the order of comparison of the data. By setting the order of operations, it is possible to have the effect of assigning weights to each input data set.

In particular, when the priority is determined according to the order of comparison of data, that is, when the change of the first data is important, after comparing the input data and the training data, the recognition result is substantially weighted, and the weighted data The recognition rate can be increased by exerting the effect of comparing and learning the input data again at and next time.

For example, the input data set A for the gas concentration a input at time t = 0 is compared with the previously learned result (reference data) to determine the gas concentration based on the similarity. For example, if a gas concentration is determined to be 30 ppm, a danger warning is issued if the input concentration value at the next measurement time t=1 falls between 11 ppm and 35 ppm, and if the gas concentration at time t = 0 is determined as, for example, in the 10 ppm category, the following The danger warning can be set only when the gas concentration measured at the measurement time t=1 falls between 17 ppm and 22 ppm. This setting can be set in the same manner as the relationship between time t=0 and time t=1 for time t=2, and multiple layers can be configured by setting in this manner for time points t> 2 or more.

Under such a setting, the gas concentration a at time t = 0 has the same effect as having a higher weight compared to the gas concentration a at time t = 1.

In other words, different from the prior art, which determined that one threshold value was determined as risk and safety when exceeded, it is possible to set various threshold values according to the characteristics of the product, and make more sophisticated judgments that additionally determine when the corresponding range is satisfied. Do.

The same effect can be obtained by sequentially calculating different data (A gas concentration, B gas concentration) input at the same time in the same manner.

According to the present invention, by configuring multiple layers for input data using a variable memory structure, it is possible to obtain an effect of giving different weights to input data at each viewpoint.

The method of implementing this point in the neuromorphic memory can exert the above weighting effect by determining reference data for the next operation according to the current operation result.

This method has the advantage of increasing accuracy and preventing errors in the result.

14 shows a method of increasing the precision in stages of computation.

When determining the inclination of a part between 0 and 90 degrees based on an image, in order to determine the inclination in units of 1 degree in a conventional single layer structure, at least 90 training data (reference data) for each inclined image data Prepared and contrasted with the image of the parts and 90 reference data, respectively.

However, if this is determined by dividing it into, for example, three steps, the same precision can be obtained with only a comparison operation for much less reference data.

That is, in step 1, if only two training data (reference data) with an inclination of 0 degrees and 90 degrees are compared, and as a result, if the rotation angle is determined to be closer to 0 degrees than 90 degrees, five reference data in increments of 10 degrees in step 2 Is prepared and compared with the image data, and as a result, if it is determined that it is the closest to 10 degrees, in step 3, the reference data within the range of 5 to 15 degrees and the input image data are compared with 11 reference data in units of 1 degree. Find reference data with high weighted similarity and determine the inclination angle.

In this case, in determining the inclination with a precision of 1 degree, there are only 18 contrasting reference data in total of 3 steps.

When the number of training data compared to the input data decreases, processing load and power consumption due to movement of the reference data can be significantly reduced.

In addition, since the data operation method according to the present invention is processed under a variable neuromorphic memory structure, it is possible to have a plurality of input data. Therefore, after calculating a weight on the input data, the weighted data is viewed as new input data and a comparison operation is performed. You can derive results after performing As a result, compared to a method having one hidden layer, the accuracy is high, and the same effect as deep learning can be obtained.

In particular, it has the advantage of being able to derive new data through fusion of multiple data and to learn new data derived from it. This method has the advantage of being able to apply complex algorithms because it can convert a single layer into a multi-layer method and perform regression operations.

In the above-described embodiment, the prediction of the size of the input data and the input timing is performed by the PSE 220, and the allocation of the memory area for the input data set and the reference data set and the storage of each data setter are performed by the PSM 210. do. The order and timing of operations for each input data set may be controlled by the PSM 210 or a separate operation controller (not shown).

In the above, the configuration of the present invention has been described in detail with reference to some embodiments, but this is only an example and the technical idea of the present invention is not limited thereto.

Those of ordinary skill in the art to which the present invention pertains will appreciate that it is possible to easily transform into other specific forms without changing the technical spirit or essential features of the present invention. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not limiting.

For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form. For example, although it has been described that the PSM and PSE are separate software modules or hardware modules in each embodiment, they may be implemented and operated as a combined soft air or hardware.

In addition, although the systems and methods of the present invention have been described in connection with specific embodiments, some or all of their components or operations may be implemented using a computer system having a general-purpose hardware architecture.

In particular, in the above embodiments, a memory in a neuromorphic semiconductor has been described as a main target, but it is a matter of course that a configuration applicable to other types of memory among the technical concepts of the present invention can be applied to other types of memory management.

Accordingly, the scope of the present invention is defined by the claims to be described later, and all changes or modified forms derived from the equivalent concept should be interpreted as being included in the scope of the present invention.

Claims (12)

A memory having a multidimensional lattice structure,
A prediction unit that predicts a degree of memory usage of each of two or more data sets and determines an order of input to the memory;
A management unit that stores, learns, and infers the data set by allocating a memory area to store the two or more data sets according to the prediction result of the prediction unit
Memory management system comprising a.
The method of claim 1, wherein the memory comprises:
A memory management system that has one of the two-dimensional and three-dimensional lattice structures and can be accessed from two or more directions according to each dimension.
The method of claim 1, wherein the prediction unit,
A memory management system to predict the type and size of data input in a predetermined time interval based on at least one of pre-stored reference information and executed application output data information.
The method of claim 1, wherein the prediction unit,
For the two or more data sets, determining the order of memory allocation in the order of the size of each predicted data set
In-memory management system.
The method of claim 1, wherein the prediction unit,
Determining a memory allocation order and reallocating a memory region based on the predicted size of each data set and the number of cells on one side of the memory region temporarily set for each data set for the two or more data sets
In-memory management system.
A memory having a multidimensional lattice structure,
A prediction unit that predicts the degree of memory usage of each of two or more data sets,
A management unit that stores, learns, and infers the data set by allocating a memory area to store the two or more data sets according to the lambda element technique according to the prediction result of the prediction unit
Memory management system comprising a.
The method of claim 6, wherein the memory,
A storage unit for storing the two or more data sets and reference data to be used for calculation,
An operation unit that calculates the data set based on the reference data
A memory management system comprising a.
The method of claim 6, wherein the management unit,
And controlling the memory to perform learning on the two or more data stored in the memory and store the learning result.
The method of claim 6, wherein the management unit,
And storing the two or more data alternately with respect to the two or more axial directions of the memory.
In the neuromorphic semiconductor memory management method,
Allocating and storing a memory area for two or more input data sets,
Performing learning on the two or more input data sets,
Storing reference data as the learning result in a new memory area,
Comparing and calculating the reference data on one or more new input data sets.
Neuromorphic memory management method comprising a.
The method of claim 10, wherein the storing comprises:
Predicting the size of each of the two or more input data sets; and
First allocating and storing a memory area for an input data set having a large size among the input data sets according to the prediction result
Neuromorphic memory management method comprising a.
The method of claim 10,
Determining reference data to be used for a subsequent calculation of another new input data set among the stored reference data according to the result of the calculation step
Neuromorphic memory management method further comprising a.

Images