Meta-model extension for machine learning equalizer

By using multiple shallow threshold expert machine learning models of equalizers and metanets for denoising in NAND flash devices, the problem of noise impact in multi-layer memory cells is solved, and low-power and efficient data reading is achieved. for mobile electronic devices.

CN119028401APending Publication Date: 2024-11-26SAMSUNG ELECTRONICS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410587617.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-23
Filing Date
2024-05-13
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The prior art is difficult to effectively reduce the noise impact in NAND flash memory devices in mobile electronic devices, especially in multi-layer memory units, resulting in uncertainty in reading information and failure of ECC operations, and conventional equalization modules require high computing power. Influence performance.

Method used

The NAND unit is denoised by the equalizer of multiple shallow threshold expert machine learning models, and parameters are generated for each threshold expert through the meta network, and appropriate threshold expert is selected based on the weak decision voltage range for data classification and denoising, generating Improved memory channel output.

Benefits of technology

It realizes a low-power and low-memory denoising solution, improves the efficiency and accuracy of ECC operation, and is suitable for various NAND flash architectures, reducing latency and memory requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028401A_ABST
    Figure CN119028401A_ABST
Patent Text Reader

Abstract

Systems and methods of the present disclosure may be used to improve an equalization module architecture for NAND unit read information. For example, embodiments of the present disclosure may provide for denoising NAND cell read information using a multi-shallow threshold expert machine learning model (MTM) equalizer. The MTM equalizer may include a plurality of shallow machine learning models. The meta-network may generate parameters for each of the shallow machine learning models such that each shallow machine learning model may be able to address a classification task (e.g., a binary classification task) corresponding to a weak decision range between two possible read information values for a given NAND cell read operation. Thus, during inference, each read sample having read values within a weak decision range may be passed through a corresponding shallow machine learning model (e.g., a corresponding threshold expert) associated with a particular weak decision range.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The following relates generally to flash denoising, and more specifically to flash denoising using a machine learning model. Background Art

[0002] Memory devices are common electronic components used to store data. NAND flash memory devices allow several bits of data to be stored in each memory cell, thereby providing improvements in manufacturing cost and performance. Memory cells in which multiple bits of data are stored may be referred to as multi-level memory cells. Multi-level memory cells divide the threshold voltage range of the memory cell into several voltage states, and the data value written to the memory cell is extracted using the memory cell voltage levels.

[0003] However, storing multiple bits per memory cell may reduce the dynamic voltage range of each voltage state, thereby making the memory cell more susceptible to noise. Compensating for such noise may require increased computing power, which may hinder performance in mobile devices. Therefore, there is a need in the art for a reliable, low-power multi-cell memory system for mobile electronic devices. Summary of the invention

[0004] A method for reading data from a memory device is described. An embodiment of the method may read a block of data from a memory device (e.g., a NAND cell); process the block of data using an equalizer having multiple shallow threshold expert machine learning models to classify each bit of the block of data using a log likelihood ratio (LLR) value; and decode the block of data based on an error correction code (ECC) encoding and decoding scheme. A metamodel may be used to generate parameters for each threshold expert machine learning model, and the threshold expert machine learning model may classify bits based on the parameters generated by the metamodel.

[0005] A method, apparatus, non-transitory computer-readable medium, and system for flash memory denoising using a machine learning model are described. One or more embodiments of the method, apparatus, non-transitory computer-readable medium, and system are configured to: receive first information for a plurality of memory cells of a memory device and second information for a memory cell in the plurality of memory cells, generate one or more threshold expert parameters for a threshold expert based on the first information, select a threshold expert from a plurality of threshold experts based on the second information, wherein each of the threshold experts corresponds to at least one of a plurality of weak decision voltage ranges for the memory cell, and generate a memory channel output based on the second information using the selected threshold expert and the one or more threshold expert parameters.

[0006] An apparatus, system, and method for flash memory denoising using a machine learning model are described. Embodiments of the apparatus, system, and method include: a memory device including a plurality of memory cells arranged in a plurality of word lines; a meta-network configured to generate one or more threshold expert parameters for a threshold expert based on first information for a word line in the plurality of word lines; and a plurality of threshold experts configured to generate a memory channel output based on second information from the memory cells in the word line and the one or more threshold expert parameters.

[0007] A non-transitory computer-readable medium, apparatus, system, and method for flash memory denoising using a machine learning model are described. One or more embodiments of the method, apparatus, non-transitory computer-readable medium, and system are configured to: receive first information for a plurality of memory cells of a memory device and second information for a memory cell in the plurality of memory cells, generate one or more threshold expert parameters for a threshold expert based on the first information, select a threshold expert from a plurality of threshold experts based on the second information, wherein each of the threshold experts corresponds to at least one of a plurality of weak decision voltage ranges for the memory cell, and generate a memory channel output based on the second information using the selected threshold expert and the one or more threshold expert parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Features of the present inventive concept will become more apparent by describing embodiments of the present inventive concept in detail with reference to the attached drawings.

[0009] Figure 1 is a block diagram illustrating an implementation of a data processing system including a memory system according to an embodiment of the inventive concept.

[0010] Figure 2 FIG. 1 is a diagram showing an embodiment according to the present invention. Figure 1Block diagram of the memory system.

[0011] Figure 3 According to the embodiment of the present invention Figure 1 Detailed block diagram of a non-volatile memory device.

[0012] Figure 4 According to the embodiment of the present invention Figure 2 Block diagram of a memory cell array.

[0013] Figure 5 According to the embodiment of the present invention Figure 4 Circuit diagram of a memory block of a memory cell array.

[0014] Figure 6 An example of an equalizer architecture according to aspects of the present disclosure is shown.

[0015] Figure 7 An example of a process for flash denoising using a machine learning model in accordance with aspects of the present disclosure is shown.

[0016] Figure 8 An example of a block diagram illustrating an implementation of a data processing system in accordance with aspects of the present disclosure is shown.

[0017] Fig. 9 An example of an equalizer architecture using multiple threshold expert machine learning models in accordance with aspects of the present disclosure is shown.

[0018] Fig.10 An example of a meta-network architecture according to aspects of the present disclosure is shown.

[0019] Fig.11 An example of a sample denoising scheme according to aspects of the present disclosure is shown. DETAILED DESCRIPTION

[0020] The present disclosure relates to systems and methods for reading data from a memory device, and more particularly to denoising data read from a memory device. Certain embodiments of the present disclosure relate to generating parameters for multiple threshold expert machine learning models and denoising NAND cell read information using an equalizer with multiple threshold expert machine learning models.

[0021] Memory devices are common electronic components used to store data. NAND flash memory devices allow several bits of data to be stored in each memory cell, providing improvements in manufacturing cost and performance. Memory cells in which multiple bits of data are stored may be referred to as multi-layer memory cells. Multi-layer memory cells divide the threshold voltage range of the memory cell into several voltage states, and the data value written to the memory cell is extracted using the memory cell voltage levels.

[0022] In order to read information from a memory device, the voltage of each cell is measured and the voltage level stored in the cell is inferred. The bit can be recovered later. There may be a trade-off between the number of voltage levels and memory reliability. The larger the number of bits per cell, the more information can be stored on the device and the more bits in each cell. Alternatively, voltages representing different levels can be packed more tightly together because a greater number of distinguishable voltages are used within the same dynamic range. Therefore, noise in cell programming or cell reading is more likely to change the voltage of one level to another voltage representing a different level, thereby showing an error when reading the cell. In other words, when reading information from a cell (e.g., a NAND cell), the read information may be corrupted by noise (e.g., depending on the characteristics of the medium), which may result in an ambiguous value of the read information (e.g., the read voltage level may be between two voltage values, where the voltage information of different values ​​of the detected channel output is likely to overlap in a weak decision voltage range). In some examples, such noise may cause an error correction code (ECC) operation to fail.

[0023] Conventional approaches for handling ECC operation failures may include utilizing an equalization module in an effort to reduce inter-symbol-interference (ISI) across NAND cells. Such an equalization module may receive sample (memory cell) characteristics (e.g., cell voltage, pillar and word line neighbors, thresholds, etc.) as input, and the equalization module may output improved soft symbol estimates (e.g., which may allow for more efficient ECC operations with improved soft symbol estimates). However, the equalization module may pass the samples through the same model (e.g., such as a neural network) for soft symbol estimation across different voltage levels. Such an approach may require a relatively complex (e.g., deep, computationally intensive, etc.) machine learning model with high latency and memory requirements, and may not provide optimal performance. Furthermore, due to the high latency / memory requirements, only limited NAND architectures may be able to support such an equalization module.

[0024] The systems and methods of the present disclosure can be used to improve the equalization module architecture for NAND cell read information. For example, embodiments of the present disclosure can provide denoising of NAND cell read information using a multi-shallow threshold expert machine learning model (MTM) equalizer. The MTM equalizer may include multiple shallow machine learning models, each of which is used to specifically solve a classification task (e.g., a binary classification task) corresponding to a weak decision range between two possible read information values ​​for a given NAND cell read operation. Therefore, during inference, each read sample with a read value within the weak decision range is passed through a corresponding shallow machine learning model (e.g., a corresponding threshold expert) associated with a specific weak decision range. Multiple threshold expert shallow machine learning models can therefore output improved soft symbol estimates for their corresponding weak decision ranges.

[0025] The meta-network can be trained to generate parameters for each threshold expert of the MTM equalizer, and the threshold expert can classify bits based on the parameters generated by the meta-network for the threshold expert. The parameters generated for the threshold expert can allow the threshold expert to handle the classification task associated with a specific weak decision range. The improved soft symbol estimate can be used to generate an improved codeword for subsequent ECC operations (for example, it can improve EEC operation efficiency). The use of multiple shallow machine learning models (each including parameters that allow the machine learning model to solve a specific binary classification task associated with a specific weak decision range) can also produce a set of machine learning models that can operate with low memory usage (for example, less memory usage than ECC) and low latency. In addition, such an equalization model can be implemented in firmware so that the equalization model can be used by various NAND flash memory architectures. In addition, because the meta-network can be trained to generate parameters for a threshold expert that can classify bits in any of a plurality of weak decision ranges, a machine learning model with multiple threshold experts can be quickly configured to work with memory cells with various weak decision ranges (for example, without having to train each threshold expert independently).

[0026] Embodiments of the inventive concept will be described more fully hereinafter with reference to the accompanying drawings.Like reference numerals may refer to like elements throughout the drawings.

[0027] It will be understood that the terms "first", "second", "third", etc. are used in this article to distinguish one element from another element, and these elements are not limited by these terms. Therefore, the "first" element in an exemplary embodiment can be described as the "second" element in another exemplary embodiment. It should be understood that the description of features or aspects in each exemplary embodiment should typically be considered to be applicable to other similar features or aspects in other exemplary embodiments, unless the context clearly indicates otherwise. As used herein, the singular forms of "one", "an", and "the" are intended to also include plural forms, unless the context clearly indicates otherwise.

[0028] In this article, as will be understood by those of ordinary skill in the art, when a value is described as being approximately equal to another value or being substantially the same or equal to another value, it will be understood that these values ​​are within the measurement error of each other, or if they are not equal to the extent that can be measured, then these values ​​are close enough to be functionally equal to each other. For example, the term "about" as used herein includes the value, and means being within an acceptable deviation range for a particular value as determined by those of ordinary skill in the art taking into account the measurement in question and the error associated with the measurement of a particular amount (i.e., the limitations of the measurement system). For example, as will be understood by those of ordinary skill in the art, "about" can mean being within one or more standard deviations. In addition, as will be understood by those of ordinary skill in the art, it should be understood that when a parameter can be described as "approximately" having a specific value in this article, according to an exemplary embodiment, the parameter can be exactly that specific value or approximately that specific value within the measurement error.

[0029] Example Memory Devices

[0030] Figure 1-5 An example memory system according to aspects of the present disclosure is shown. In a memory system (e.g., such as a solid-state drive (SSD) system), a memory controller is connected to several NAND channels in parallel to achieve high data throughput. The memory controller includes a signal processing and ECC engine that decodes data from the NAND and reliably retrieves the stored data. As described herein, Figure 1-5 An example memory system may include a memory controller having multiple NAND denoising threshold experts (e.g., a memory controller having an equalizer including multiple shallow threshold expert machine learning models).

[0031] Figure 1 is a block diagram illustrating an implementation of a data processing system including a memory system according to an embodiment of the inventive concept.

[0032] refer to Figure 1 , the data processing system 10 may include a host 100 and a memory system 200 . Figure 1The memory system 200 shown in FIG. 2 can be utilized in various systems including data processing functions. The various systems can be various devices including, for example, mobile devices such as smart phones or tablet computers. However, the various devices are not limited thereto.

[0033] The memory system 200 may include various types of memory devices. In this document, embodiments of the inventive concept will be described as including a memory device as a non-volatile memory. However, embodiments are not limited thereto. For example, the memory system 200 may include a memory device as a volatile memory.

[0034] According to an embodiment, the memory system 200 may include a non-volatile memory device, such as, for example, a read-only memory (ROM), a magnetic disk, an optical disk, a flash memory, etc. The flash memory may be a memory that stores data according to a change in the threshold voltage of a metal oxide semiconductor field effect transistor (MOSFET), and may include, for example, NAND and NOR flash memories. The memory system 200 may be implemented using a memory card including a non-volatile memory device, such as, for example, an embedded multimedia card (eMMC), a secure digital (SD) card, a micro SD card, or a universal flash memory (UFS), or the memory system 200 may be implemented using, for example, an SSD including a non-volatile memory device. In this document, the configuration and operation of the memory system 200 will be described under the assumption that the memory system 200 is a non-volatile memory system. However, the memory system 200 is not limited thereto. The host 100 may include, for example, a system on chip (SoC) application processor (AP) installed on a mobile device or a central processing unit (CPU) included in a computer system.

[0035] The processor is an intelligent hardware device (e.g., a general processing component, a digital signal processor (DSP), a CPU, a graphics processing unit (GPU), a microcontroller, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic device, a discrete gate or transistor logic component, a discrete hardware component, or any combination thereof). In some cases, the processor is configured to operate the memory array using a memory controller. In other cases, the memory controller is integrated into the processor. In some cases, the processor is configured to execute computer-readable instructions stored in the memory to perform various functions. In some embodiments, the processor includes dedicated components for modem processing, baseband processing, digital signal processing, or transmission processing.

[0036] As described above, the host 100 may include the AP 110. The AP 110 may include various intellectual property (IP) blocks. For example, the AP 110 may include a memory device driver 111 that controls the memory system 200. The host 100 may communicate with the memory system 200 to send commands related to memory operations and receive confirmation commands in response to the sent commands. The host 100 may also communicate with the memory system 200 regarding information tables related to memory operations.

[0037] The memory system 200 may include, for example, a memory controller 210 and a memory device 220. The memory controller 210 may receive commands related to memory operations from the host 100, generate internal commands and internal clock signals using the received commands, and provide the internal commands and internal clock signals to the memory device 220. The memory device 220 may store write data in a memory cell array in response to an internal command, or may provide read data to the memory controller 210 in response to an internal command. In some examples, the memory device 220 may be referred to as a channel. The term channel is used because write and / or send operations may go to and / or pass through a channel. In some cases, when information is read, the information may be damaged by noise depending on the characteristics of the medium.

[0038] As described herein, the memory controller 210 may include multiple threshold experts (e.g., the memory controller 210 may use an equalizer including multiple shallow threshold expert machine learning models to denoise the memory device 220 (e.g., NAND) read information). For example, the memory controller 210 may include an MTM equalizer that includes multiple threshold experts, each of which is trained to specifically solve the NAND read value classification task. In some examples, the meta-network can be used to generate parameters for each of the multiple threshold experts in the memory controller 210. The memory controller 210 is herein (e.g., reference Figure 8 ) describes examples of corresponding elements or includes various aspects thereof.

[0039] The memory device 220 includes a memory cell array that retains data stored therein (even when the memory device 220 is not powered on). The memory cell array may include, for example, NAND or NOR flash memory, magnetoresistive random access memory (MRAM), resistive random access memory (RRAM), ferroelectric random access memory (FRAM), or phase change memory (PCM) as memory cells. For example, when the memory cell array includes a NAND flash memory, the memory cell array may include a plurality of blocks and a plurality of pages. Data may be programmed and read in units of pages, and data may be erased in units of blocks. Examples of memory blocks included in the memory cell array are shown in FIG. Figure 4 Shown in.

[0040] Figure 2 FIG. 1 is a diagram showing an embodiment according to the present invention. Figure 1 Block diagram of the memory system.

[0041] refer to Figure 2 , the memory system 200 includes a memory device 220 and a memory controller 210. The memory controller 210 may also be referred to as a controller circuit in this document. The memory device 220 may perform a write operation, a read operation, or an erase operation under the control of the memory controller 210.

[0042] The memory controller 210 may control the memory device 220 depending on a request received from the host 100 or an internally specified schedule. The memory controller 210 may include a controller core 211, an internal memory 214, a host interface block 215, and a memory interface block 216. The memory controller 210 may further include a device information storage 217 configured to provide the first device information DI1 to the host interface block 215 and provide the second device information DI2 to the controller core 211.

[0043] In some embodiments, the controller core 211 may include an encoder 212 and an equalizer 213. In some embodiments, the equalizer 213 may be an example of, may include aspects of, or may implement the following: a neural network 600, a plurality of threshold experts 810, a plurality of threshold experts 905, and a plurality of threshold experts 1120. As described herein, the encoder 212 may perform an ECC operation on the read information (e.g., a cell voltage, a same-pillar neighbor voltage, etc.). In the event of a failure of the ECC operation, the read information may be passed to the equalizer 213, or the equalizer 213 may read new read information from the memory device 220.

[0044] According to the techniques described herein, the equalizer 213 can use multiple shallow threshold expert machine learning networks, use threshold experts selected for read information (e.g., threshold experts trained for classification operations in a weak decision range corresponding to the read information), and generate improved memory channel outputs based on the read information. That is, an initial memory channel output can be generated, and if the ECC operation fails, the equalizer 213 can generate an improved memory channel output that represents a better estimate of the original stored information in one or more memory cells. Therefore, the controller core 211 can generate an updated codeword and perform a successful ECC operation based on the estimated channel output (e.g., based on soft symbol estimates from one or more threshold experts from the equalizer 213).

[0045] The controller core 211 may include a memory control core, which may control and access the memory device 220 depending on a request received from the host 100 or an internally specified schedule. The memory control core may manage and execute various metadata and codes for managing or operating the memory system 200. In some embodiments, the controller core 211 may include a memory control core and a machine learning core, and each of these cores may be implemented by one or more processors. In some examples where the memory controller 210 includes a machine learning core, the machine learning core may be used to perform reference Figure 6 and Figure 8-Figure 11 Training and inference of a threshold expert network (eg, a shallow threshold expert machine learning network) described herein. For example, such a threshold expert network can be designed and trained to perform equalization or denoising operations on read information read from the memory device 220.

[0046] The internal memory 214 may be used as, for example, a system memory used by the controller core 211, a cache memory storing data of the memory device 220, or a buffer memory temporarily storing data between the host 100 and the memory device 220. The internal memory 214 may store a mapping table (MT) indicating a relationship between a logical address assigned to the memory system 200 and a physical address of the memory device 220. The internal memory 214 may include, for example, a DRAM or an SRAM.

[0047] The host interface block 215 may include components for communicating with the host 100, such as, for example, a physical block. The memory interface block 216 may include components for communicating with the memory device 220, such as, for example, a physical block.

[0048] Hereinafter, the operation of the memory system 200 over time will be described. When power is supplied to the memory system 200 , the memory system 200 may perform initialization with the host 100 .

[0049] The host interface block 215 may provide the memory control core with a first request REQ1 received from the host 100. The first request REQ1 may include a command (eg, a read command or a write command) and a logical address. The memory control core may convert the first request REQ1 into a second request REQ2 suitable for the memory device 220.

[0050] For example, the memory control core may convert the format of the command. The memory control core may obtain address information by referring to the mapping table MT stored in the internal memory 214. The memory control core may convert the logical address into the physical address of the memory device 220 by using the address information. The memory control core may provide the memory interface block 216 with a second request REQ2 suitable for the memory device 220.

[0051] The memory interface block 216 may register the second request REQ2 from the memory control core at the queue. The memory interface block 216 may send the request first registered at the queue to the memory device 220 as the third request REQ3.

[0052] When the first request REQ1 is a write request, the host interface block 215 may write data received from the host 100 to the internal memory 214. When the third request REQ3 is a write request, the memory interface block 216 may transmit data stored in the internal memory 214 to the memory device 220.

[0053] When the data is completely written, the memory device 220 may send a third response RESP3 to the memory interface block 216. In response to the third response RESP3, the memory interface block 216 may provide the memory control core with a second response RESP2 indicating that the data is completely written.

[0054] After the data is stored in the internal memory 214 or after the second response RESP2 is received, the memory control core may send a first response RESP1 indicating request completion to the host 100 through the host interface block 215 .

[0055] When the first request REQ1 is a read request, the read request may be sent to the memory device 220 through the second request REQ2 and the third request REQ3. The memory interface block 216 may store the data received from the memory device 220 in the internal memory 214. When the data is completely sent, the memory device 220 may send a third response RESP3 to the memory interface block 216.

[0056] When the third response RESP3 is received, the memory interface block 216 may provide the memory control core with a second response RESP2 indicating that the data is completely stored. When the second response RESP2 is received, the memory control core may send the first response RESP1 to the host 100 through the host interface block 215 .

[0057] The host interface block 215 may transmit data stored in the internal memory 214 to the host 100. In an embodiment, in a case where data corresponding to the first request REQ1 is stored in the internal memory 214, transmission of the second request REQ2 and the third request REQ3 may be omitted.

[0058] The memory device 220 may also transmit first serial peripheral interface information SPI1 to the memory interface block 216. The memory interface block 216 may transmit second serial peripheral interface information SPI2 to the controller core 211.

[0059] Figure 3 According to the embodiment of the present invention Figure 1 A detailed block diagram of the non-volatile memory device 220 is shown in FIG. Figure 3 , the memory device 220 may include, for example, a memory cell array 221 , a control logic 222 , a voltage generating unit 223 , a row decoder 224 , and a page buffer 225 .

[0060] The memory cell array 221 may be connected to one or more string selection lines (SSL), a plurality of word lines WL, one or more ground selection lines (GSL), and a plurality of bit lines BL. The memory cell array 221 may include a plurality of memory cells arranged at intersections between a plurality of word lines and a plurality of bit lines.

[0061] The control logic 222 may receive a command CMD (e.g., an internal command) and an address ADD from the memory controller 210, and receive a control signal CTRL for controlling various functional blocks within the memory device 220 from the memory controller 210. The control logic 222 may output various control signals for writing data to or reading data from the memory cell array 221 based on the command CMD, the address ADD, and the control signal CTRL. In this manner, the control logic 222 may control the overall operation of the memory device 220.

[0062] Various control signals output by the control logic 222 may be provided to the voltage generation unit 223, the row decoder 224, and the page buffer 225. For example, the control logic 222 may provide a voltage control signal CTRL_vol to the voltage generation unit 223, a row address X-ADD to the row decoder 224, and a column address Y-ADD to the page buffer 225.

[0063] The voltage generating unit 223 may generate various voltages for performing programming, reading and erasing operations on the memory cell array 221 based on the voltage control signal CTRL_vol. For example, the voltage generating unit 223 may generate a first driving voltage VWL for driving a plurality of word lines WL, a second driving voltage VSSL for driving a plurality of string selection lines SSL, and a third driving voltage VGSL for driving a plurality of ground selection lines GSL. In this case, the first driving voltage VWL may be a programming voltage (e.g., a write voltage), a read voltage, an erase voltage, a pass voltage, or a programming verification voltage. In addition, the second driving voltage VSSL may be a string selection voltage (e.g., a turn-on voltage or a turn-off voltage). In addition, the third driving voltage VGSL may be a ground selection voltage (e.g., a turn-on voltage or a turn-off voltage).

[0064] The row decoder 224 may be connected to the memory cell array 221 through a plurality of word lines, and may activate a portion of the plurality of word lines in response to a row address X-ADD received from the control logic 222. For example, in a read operation, the row decoder 224 may apply a read voltage to a selected word line and a pass voltage to unselected word lines.

[0065] In a program operation, row decoder 224 may apply a program voltage to a selected word line and a pass voltage to unselected word lines. In an embodiment, in at least one of a plurality of program loops, row decoder 224 may apply a program voltage to the selected word line and additionally selected word lines.

[0066] The page buffer 225 may be connected to the memory cell array 221 through a plurality of bit lines BL. For example, in a read operation, the page buffer 225 may operate as a sense amplifier that outputs data stored in the memory cell array 221. Alternatively, in a program operation, the page buffer 225 may operate as a write driver that writes desired data to the memory cell array 221.

[0067] Figure 4 and 5 An example is shown in which the memory system 200 is implemented using a three-dimensional flash memory. The three-dimensional flash memory may include a three-dimensional (e.g., vertical) NAND (e.g., VNAND) memory cell. The implementation of the memory cell array 221 including the three-dimensional memory cell is described below. Each of the memory cells described below may be a NAND memory cell.

[0068] Figure 4 According to the embodiment of the present invention Figure 2 A block diagram of a memory cell array 221 is shown.

[0069] refer to Figure 4 , the memory cell array 221 according to the embodiment includes a plurality of memory blocks BLK1 to BLKz. Each of the memory blocks BLK1 to BLKz has a three-dimensional structure (e.g., a vertical structure). For example, each of the memory blocks BLK1 to BLKz may include a structure extending in a first direction to a third direction. For example, each of the memory blocks BLK1 to BLKz may include a plurality of NAND strings extending in a second direction. The plurality of NAND strings may be provided, for example, in a first direction to a third direction.

[0070] Each of the NAND strings is connected to a bit line BL, a string selection line SSL, a ground selection line GSL, a word line WL, and a common source line CSL. That is, each of the memory blocks BLK1 to BLKz may be connected to a plurality of bit lines BL, a plurality of string selection lines SSL, a plurality of ground selection lines GSL, a plurality of word lines WL, and a common source line CSL. Figure 5 The memory blocks BLK1 to BLKz are described in more detail.

[0071] Figure 5 is a circuit diagram of a memory block BLKi according to an embodiment of the inventive concept. Figure 5 Shows Figure 4 An example of one of the memory blocks BLK1 to BLKz in the memory cell array 221. Figure 5 An example with 6 word lines and 6 memory cells is shown, but this is merely an illustration and any number of word lines and memory cells may be used.

[0072] The memory block BLKi may include a plurality of cell strings CS11 to CS41 and CS12 to CS42. The plurality of cell strings CS11 to CS41 and CS12 to CS42 may be arranged in the column direction and the row direction to form columns and rows. Each of the cell strings CS11 to CS41 and CS12 to CS42 may include a ground selection transistor GST, memory cells MC1 to MC6, and a string selection transistor SST. The ground selection transistor GST, memory cells MC1 to MC6, and string selection transistor SST included in each of the cell strings CS11 to CS41 and CS12 to CS42 may be stacked in a height direction substantially perpendicular to the substrate.

[0073] The columns of the plurality of cell strings CS11 to CS41 and CS12 to CS42 may be connected to different string selection lines SSL1 to SSL4, respectively. For example, the string selection transistors SST of the cell strings CS11 and CS12 may be connected in common to the string selection line SSL1. The string selection transistors SST of the cell strings CS21 and CS22 may be connected in common to the string selection line SSL2. The string selection transistors SST of the cell strings CS31 and CS32 may be connected in common to the string selection line SSL3. The string selection transistors SST of the cell strings CS41 and CS42 may be connected in common to the string selection line SSL4.

[0074] The rows of multiple cell strings CS11 to CS41 and CS12 to CS42 may be connected to different bit lines BL1 and BL2, respectively. For example, the string selection transistors SST of the cell strings CS11 to CS41 may be commonly connected to the bit line BL1. The string selection transistors SST of the cell strings CS12 to CS42 may be commonly connected to the bit line BL2.

[0075] Columns of a plurality of cell strings CS11 to CS41 and CS12 to CS42 may be connected to different ground selection lines GSL1 to GSL4, respectively. For example, the ground selection transistors GST of the cell strings CS11 and CS12 may be connected in common to the ground selection line GSL1. The ground selection transistors GST of the cell strings CS21 and CS22 may be connected in common to the ground selection line GSL2. The ground selection transistors GST of the cell strings CS31 and CS32 may be connected in common to the ground selection line GSL3. The ground selection transistors GST of the cell strings CS41 and CS42 may be connected in common to the ground selection line GSL4.

[0076] Memory cells arranged at the same height from the substrate (or ground selection transistor GST) can be connected to a single word line in common, and memory cells arranged at different heights from the substrate can be connected to different word lines WL1 to WL6, respectively. For example, memory cell MC1 can be connected to word line WL1 in common. Memory cell MC2 can be connected to word line WL2 in common. Memory cell MC3 can be connected to word line WL3 in common. Memory cell MC4 can be connected to word line WL4 in common. Memory cell MC5 can be connected to word line WL5 in common. Memory cell MC6 can be connected to word line WL6 in common. The ground selection transistors GST of cell strings CS11 to CS41 and CS12 to CS42 can be connected to a common source line CSL in common.

[0077] Denoising

[0078] Figure 6 An example of an equalizer architecture according to aspects of the present disclosure is shown. The example shown includes a neural network 600, read information 605, and symbol estimation 610. The neural network 600 may be included in the reference Figure 2 The read information 605 is referenced to the equalizer 213 described herein or implemented via the equalizer 213. Fig. 9 Examples of corresponding elements described herein or including aspects thereof. Symbol estimation 610 is a reference Fig. 9 Examples of corresponding elements are described or include aspects thereof.

[0079] Neural network 600 is a computer algorithm that can learn specific patterns without being explicitly programmed, but through iteration on known data. Neural network 600 may refer to a cognitive model that includes input nodes, hidden nodes, and output nodes. Nodes in the network may have an activation function that calculates whether the node is activated based on the output of previous nodes. The training system may include providing values ​​for inputs and modifying edge weights and activation functions (algorithmically or randomly) until the results are very close to the set of desired outputs.

[0080] Artificial neural network 600 is a hardware or software component including multiple connected nodes (i.e., artificial neurons), which roughly correspond to neurons in the human brain. Each connection or edge sends a signal from one node to another node (such as a physical synapse in the brain). When a node receives a signal, it processes the signal and then sends the processed signal to other connected nodes. In some cases, the signal between nodes includes real numbers, and the output of each node is calculated by a function of the sum of its inputs. Each node and edge is associated with one or more node weights that determine how the signal is processed and sent. During the training process, these weights are adjusted to improve the accuracy of the results (i.e., by minimizing the loss function that corresponds to the difference between the current result and the target result in some way). The weight of the edge increases or decreases the strength of the signal sent between the nodes. In some cases, the node has a threshold value, below which the signal will not be sent at all. In some examples, the nodes are aggregated into layers. Different layers perform different transformations on their inputs. The initial layer is called the input layer, and the last layer is called the output layer. In some cases, the signal passes through a specific layer many times.

[0081] As described herein, an equalizer (e.g., an equalization module) is used to minimize inter-symbol interference (ISI) across NAND cells. The equalization module receives sample (memory cell) features. For example, sample features may include cell voltage, pillar and word line neighbors, thresholds, etc. Some equalization modules assume that all samples pass through the same neural network (e.g., such as neural network 600). In other words, neural network 600 can analyze all samples and perform complex classification operations to classify various samples into one of several possible voltage states. The same features are fed to neural network 600 (e.g., resulting in a total of 16 probabilities associated with each level). For example, Figure 6 The equalizer architecture shown in the example of can be used for 3-bit per cell (e.g., triple-level cell (TLC)) flash memory.

[0082] In some examples, a neural network (such as neural network 600) may include a complex machine learning model (e.g., associated with high latency and relatively many memory parameters that may be used for complex classification operations). Due to the high latency and relatively many memory parameters, a particular NAND architecture is used to support an equalization module. Some NAND architectures may not be able to support certain equalization modules (e.g., due to the complexity of certain neural network architectures for classifying various samples into one of several possible voltage states).

[0083] exist Figure 6In an example of , the equalizer architecture includes a neural network 600 that outputs soft symbol estimates 610 based on input read information 605. The read information 605 may include cell voltage information 615, same column neighbor voltage information 620, same word line neighbor voltage information 625, word line hard decision (HD) threshold (e.g., all HD thresholds) information 630, and word line index information 635, etc. In some examples, all samples (e.g., all read information 605) may pass through the neural network 600 (e.g., and the neural network 600 outputs soft symbol estimates 610 for all read information 605).

[0084] In some examples, a scheme for error correction can be based on multiple shallow threshold expert models (e.g., threshold experts) rather than a single neural network 600. Such a scheme can be used for error correction in the equalization stage (e.g., before ECC is applied). A class of equalizers that use multiple shallow threshold expert models can be referred to as multi-threshold expert models (MTEMs). In some v-NAND devices, each cell can store N (e.g., 3 or 4) bits of information using a voltage. In the denoising or equalization stage, the voltage stored in the cell can be mapped to 2 corresponding to all possible combinations of the N bits. N One of a number of possible values ​​(e.g., levels).

[0085] By considering the case where the voltage of a cell is close to the decision boundary between two levels, the threshold expert model can split the denoising problem into a single-bit decision problem. Cells whose voltage is close to the decision boundary between two levels can be referred to as weak cells. The soft symbol estimate 610 output by the neural network 600 can include an output for a weak cell. The output for a weak cell can be a single LLR value: where X is a vector of features measured for each cell (e.g., read information 605), and F w is a function with a small number of parameters (w). For each unit of N bits, there can be 2 N Levels and 2 N -1 threshold model.

[0086] During the denoising phase, the equalizer architecture (e.g., or another error correction device) can scan all cells (e.g., in parallel) and can identify weak cells, find a corresponding threshold model for each weak cell, and apply the threshold model to the weak cell. However, in some examples, the equalizer architecture may have two conflicting performance measures: accuracy and complexity cost. Accuracy can refer to the error correction capability of the equalizer architecture, while complexity cost can refer to the cost of using the model (e.g., in terms of waiting time or gate count). In order to improve the model's ability to correct bit flip errors, it may be appropriate for the read information 605 to include larger feature vectors and for each threshold expert model to include more parameters. However, larger models may result in increased computation time (e.g., in firmware) or larger gate counts (e.g., in hardware). Therefore, improving accuracy may result in higher complexity costs.

[0087] Shallow models may include a small number of features and may accept fewer feature inputs. Therefore, shallow models may prefer to reduce complexity costs rather than increase accuracy. An example architecture of a shallow threshold expert model selected for implementation in a mobile controller may be a linear function (F w ): target cell voltage, upper neighbor along pillar voltage, lower neighbor along pillar voltage, and threshold reference voltage (e.g., voltage used for hard decision without equalizer). Each threshold expert model may have five parameters (e.g., 4 weights and bias terms multiplied by input features). Thus, in total, there may be 5×(2 N -1) parameters.

[0088] A model that includes a shallow threshold expert model that accepts a feature input may be referred to as a linear MTEM. Linear MTEM may be implemented in hardware or firmware (e.g., depending on the balance between CPU latency and gate cost). In some examples, training a model with multiple threshold expert models may include selecting a single operating point, such as the number of program-erase (PE) cycles and the hold time. During training, random (but known) bits may be written to the chip, brought to the operating point, and the features of the bits may be measured. The data may then be split into corresponding threshold models (e.g., for weak cells), and a single threshold model may be trained (e.g., using machine learning techniques) to correctly estimate the LLRs.

[0089] Despite its simplicity, a model that includes multiple threshold expert models can achieve near-optimal accuracy. For example, the model may experience an accuracy gap or drop of up to 5% in corrected bits compared to a high-performance model. The accuracy gap can be measured at the same operating point used for training (e.g., on a different chip). However, in some examples, although a model that includes multiple threshold expert models can achieve high accuracy, the performance of the model may be lower at operating points far from the training operating point. Larger models can take advantage of additional features (e.g., word line index, PE cycle count, etc.) to adjust for changes in operating points. However, increasing the size of the threshold expert models may increase the complexity cost associated with these models.

[0090] In some embodiments, instead of increasing the size of the threshold expert, a meta-network can be used to generate parameters for the threshold expert model to allow the accuracy of the threshold expert model to be increased (e.g., at various operating points) while keeping the complexity cost of the threshold expert model low. For example, the meta-network can be implemented in a neural network including multiple threshold expert models to share performance costs while maintaining near-optimal accuracy for multiple (e.g., all) operating points.

[0091] The meta-network may use wordline features (e.g., features or characteristics of wordlines in a memory) to generate parameters for each threshold expert model in a set of threshold expert models. One example of a wordline feature is a set of block PE cycles corresponding to the number of PE cycles of a target block. Another example of a wordline feature is a (e.g., within a column) tier index corresponding to a page index divided by the number of pages in a layer (e.g., without remainder). Another example of a wordline feature is an SSL index corresponding to a layer of a target page (e.g., a layer in which the target page is embedded). Another example of a wordline feature is an extremely strong error ratio corresponding to a percentage of cells in a sector with extremely strong soft decision symbols. Another example of a wordline feature is a normal strong error ratio corresponding to a percentage of cells in a sector with normal strong soft decision symbols. Another example of a wordline feature is a normal weak error ratio corresponding to a percentage of cells in a sector with normal weak soft decision symbols. Another example of a wordline feature is an extremely weak error ratio corresponding to a percentage of cells in a sector with extremely weak soft decision symbols. Another example of a word line characteristic is an HD threshold voltage (eg, all HD threshold voltages) corresponding to a reference voltage used to read HD data from a cell.Another example of a word line characteristic is a soft decision bit histogram.

[0092] Denoising using MTM

[0093] Figure 7Examples of processes for flash memory denoising using machine learning models according to aspects of the present disclosure are shown. In some examples, the operations are performed by a system including a processor that executes a set of codes to control functional elements of the device. Additionally or alternatively, certain processes are performed using dedicated hardware. In general, the operations are performed according to the methods and processes described according to aspects of the present disclosure. In some cases, the operations described herein are composed of various sub-steps or performed in conjunction with other operations.

[0094] Figure 7 The process shown for flash memory denoising can employ machine learning techniques to minimize ISI across NAND cells and reduce bit error rate (BER) and fine frame error rate (FER). A meta-network can be used to adapt a shallow threshold expert to changing word lines and operating points (e.g., instead of adjusting the model for a single operating point, all word lines share the same model, and the operating point is determined by the number of PE cycles and hold time). Using a meta-network to generate parameters for the threshold expert can allow for improved BER reduction and improved LLR calibration. In addition, the meta-network can be trained to adjust the threshold expert parameters across multiple operating points (e.g., instead of operating the threshold expert with fixed parameters and training the threshold expert on a single operating point).

[0095] At operation 700, the system receives first information for a plurality of memory cells of a memory device and second information for a memory cell in the plurality of memory cells. In some cases, the operation of this step involves as described in reference Figure 8 The described meta-network, voltage detector, or both, or may be performed thereby.

[0096] At operation 705, the system generates one or more threshold expert parameters for the threshold expert based on the first information. In some cases, the operation of this step involves Figure 8 A meta-network described, or capable of being executed thereby.

[0097] At operation 710, the system selects a threshold expert from a plurality of threshold experts based on the second information, wherein each of the threshold experts corresponds to at least one weak decision voltage range of a plurality of weak decision voltage ranges of the memory cell. In some cases, the operation of this step involves as described in reference Figure 8 The selected component described, or can be executed thereby.

[0098] At operation 715, the system generates a memory channel output based on the second information using the selected threshold expert and one or more threshold expert parameters. In some cases, the operation of this step involves as described in reference Figure 8 The threshold value expert described, or can be performed by it.

[0099] Figure 8 An example of a block diagram illustrating an implementation of a data processing system according to aspects of the present disclosure is shown. The example shown includes a host device 100 and a memory system 200. The memory system 200 includes a memory controller 210 and a memory device 220. The memory controller 210 includes a voltage detector 800, a selection component 805, a plurality of threshold experts 810, an error correction component 815, and a meta-network 820. The host device 100, the memory device 220, and the memory controller 210 are reference Figure 1 and Figure 2 Examples of corresponding elements described herein or including aspects thereof. Multiple threshold experts 810 may be included in reference Figure 2 The described equalizer 213 is implemented in or via it.

[0100] Examples of memory system 200 include random access memory (RAM), read-only memory (ROM), or hard disk. Examples of memory system 200 include solid-state memory and hard disk drive. In some examples, the memory is used to store computer-readable, computer-executable software, which includes instructions that cause the processor to perform various functions described herein when executed. In some cases, the memory includes a basic input and output system (BIOS), etc., which controls basic hardware or software operations, such as interactions with peripheral components or devices. In some cases, the memory controller 210 operates the memory unit. For example, the memory controller 210 may include a row decoder, a column decoder, or both. In some cases, the memory unit in the memory stores information in the form of a logical state.

[0101] In some embodiments, an example of the memory system 200 includes flash memory. Flash memory is an electronic (solid-state) non-volatile computer storage medium that can be electrically erased and reprogrammed. The two main types of flash memory are named after NAND and NOR logic gates. Each flash memory cell exhibits internal characteristics similar to the corresponding gate. Where EPROM must be completely erased before being rewritten, NAND-type flash memory can be written and read in blocks (or pages), which are usually much smaller than the entire device. NOR-type flash memory allows a single machine word (byte) to be written to an erased location or read independently. The NAND type operates primarily in memory cards, USB flash drives, (those produced in 2009 or later) solid-state drives, and similar products for general storage and transfer of data. NAND or NOR flash memory is also commonly used to store configuration data in various digital products, a task that was previously possible through EEPROM or battery-powered static RAM. A key disadvantage of flash memory is that it can only withstand a relatively small number of write cycles in a particular block. Example applications for both types of flash memory include personal computers, PDAs, digital audio players, digital cameras, mobile phones, synthesizers, video games, scientific instruments, industrial robots, and medical electronics.

[0102] In addition to being nonvolatile, flash memory offers fast read access times, although not as fast as static RAM or ROM. Its resistance to mechanical shock helps explain its popularity over hard drives in portable devices, as does its high endurance, ability to withstand high voltages, high temperatures, and water immersion, among other things. Although flash memory is technically a type of EEPROM, the term "EEPROM" is often used to refer specifically to non-flash EEPROMs that can be erased in small blocks (usually bytes). Because erase cycles are slower, the larger block sizes used in flash erase give it a significant speed advantage over non-flash EEPROMs when writing large amounts of data. Flash memory costs much less than byte-programmable EEPROMs and has become the dominant memory type in situations where systems require large amounts of non-volatile solid-state storage.

[0103] Flash memory stores information in an array of memory cells made of floating gate transistors. In a single-level cell (SLC) device, each cell stores only one bit of information. In a multi-level cell (MLC) device (including a TLC device), each cell can store more than one bit per cell. The floating gate can be conductive or non-conductive. In NOR flash memory, one end of each cell is directly connected to the ground, and the other end is directly connected to the bit line. This arrangement is called "NOR flash memory" because it acts like a NOR gate: when one of the word lines is brought high, the corresponding storage transistor is used to pull the output bit line low. NAND flash memory also uses floating gate transistors, but it is connected in a manner similar to the NAND gate: several transistors are connected in series, and the bit line will be pulled low only when all word lines are pulled high. These several transistors connected in series are then connected to a NOR-type bit line array via some additional transistors in the same way as the single transistors in NOR flash memory are linked. Compared to NOR flash memory, an additional addressing level is added by replacing a single transistor with a series link group.

[0104] exist Figure 8 In the example of , the equalizer architecture of the memory controller 210 includes a plurality of threshold experts 810, each of which outputs a soft symbol estimate (e.g., an estimate of a channel output of a generated memory cell) to an error correction component 815 based on read information input from the memory device 220 (e.g., for the error correction component 815 to perform an ECC operation on the output soft symbol estimate). For example, the voltage detector 800 can detect read information (e.g., voltage information) from a memory cell, and the selection component 805 can select a threshold expert from the plurality of threshold experts 810 to perform a classification operation on the read information. Figure 8 In the example of , a sample (e.g., a read information sample, a detected voltage information sample, etc.) can be selected by a specific threshold expert in a plurality of threshold experts 810, wherein the specific threshold expert for each sample is selected by a selection component 805 based on the sample (e.g., based on the weak decision range in which the sample is located, as described in this article, for example, with reference to Fig.11 further described).

[0105] Thus, each of the plurality of threshold experts 810 may include a shallow machine learning model (e.g., a shallow threshold expert machine learning model) that may be associated with low latency and relatively few memory parameters. That is, the shallow threshold experts described herein may perform a binary classification operation for a specific weak decision range (e.g., as described herein, for example, with reference to Figure 9-11600). Such a classification operation may be less complex relative to a classification operation for several possible voltage states (e.g., as performed by a more complex machine learning model (such as, for example, neural network 600). Therefore, due to reduced latency and memory parameters, various NAND architectures may support such a balanced architecture that implements multiple threshold experts 810 (e.g., multiple shallow threshold expert machine learning models that each perform a respective classification operation between two corresponding voltage states).

[0106] According to some embodiments, the host device 100 may be configured to send a read command to the memory device 220 (e.g., where the read information is read by the memory controller 210 based on or in response to the read command). According to some embodiments, the host device 100 may be configured to receive a response from the memory controller 210 based at least in part on the estimated channel output.

[0107] According to some embodiments, meta-network 820 receives first information for a plurality of memory cells of memory device 220, and meta-network 820 generates one or more threshold expert parameters for a threshold expert based on the first information. In some examples, the plurality of memory cells include a word line or a subset of word lines of memory device 220. In some examples, the first information includes a number of PE cycles, a layer of a target page, a layer index, a very strong error ratio, a normal strong error ratio, a normal weak error ratio, a very weak error ratio, one or more threshold voltages, or any combination thereof. In some examples, the one or more threshold expert parameters include a weight of an upper cell voltage feature, a weight of a target cell voltage feature, a weight of a lower cell voltage feature, a weight of a threshold voltage feature, a bias term value, or any combination thereof. In some examples, meta-network 820 generates threshold expert parameters for each of the plurality of threshold experts.

[0108] According to some embodiments, the voltage detector 800 receives second information (e.g., read information) for a memory cell among a plurality of memory cells of the memory device 220. In some examples, the second information includes one or more of the following: voltage information of the memory cell, voltage information of cells in a neighboring word line of the memory cell, voltage of cells in the same word line as the memory cell, one or more HD thresholds of the word line, one or more SD thresholds of the word line, and a word line index. In some examples, the second information includes an upper cell voltage, a target cell voltage, a lower cell voltage, a threshold voltage, or any combination thereof. According to some embodiments, the voltage detector 800 may be configured to detect voltage information from a memory cell.

[0109] According to some embodiments, the selection component 805 selects a threshold expert from a plurality of threshold experts based on the second information, wherein each of the threshold experts corresponds to at least one weak decision voltage range of a plurality of weak decision voltage ranges of the memory cell. In some examples, the selection component 805 determines that the voltage value of the second information is within at least one weak decision voltage range corresponding to the threshold expert, wherein the threshold expert is selected based on the determination. In some examples, each of the weak decision voltage ranges includes an area between two hard decision thresholds where voltage information of different values ​​of the channel output is likely to overlap. In some examples, the number of threshold experts is one less than the number of hard decision thresholds. In some examples, the selection component 805 sends read information to the selected threshold expert. In some examples, the weak decision voltage ranges are spaced apart from each other by a set of strong decision voltage ranges, and wherein each of the strong decision voltage ranges corresponds to a value of the channel output.

[0110] According to some embodiments, the selection component 805 can be configured to select a threshold expert from a plurality of threshold experts 810 based on the detected voltage information. In some examples, the selection component 805 is further configured to determine that the detected voltage information corresponds to a weak decision voltage range in a set of weak decision voltage ranges, wherein the threshold expert is selected based on the determination. The selection component 805 is a reference Fig. 9 Examples of corresponding elements are described or include aspects thereof.

[0111] Multiple threshold experts 810 are reference Fig. 9 and Fig.11 Examples of corresponding elements described herein or including aspects thereof. According to some embodiments, a selected threshold expert among the plurality of threshold experts 810 generates a memory channel output based on the second information and one or more threshold expert parameters. For example, the selected threshold expert generates an improved memory channel output based on the read information. In some examples, each of the threshold experts is configured with parameters for processing the read information corresponding to a specific weak decision voltage range. In some examples, the plurality of threshold experts 810 generates LLRs based on the improved memory channel output.

[0112] According to some embodiments, a plurality of threshold experts 810 may be configured to estimate a channel output of a memory cell based on detected voltage information, wherein each of the threshold experts corresponds to a different weak decision voltage range in a plurality of weak decision voltage ranges. In some examples, each threshold expert in the set of threshold experts is further configured to generate an output value indicating whether the detected voltage information corresponds to a first value of the channel output or a second value of the channel output. In some examples, each of the threshold experts is trained to operate on read information corresponding to a particular weak decision voltage range (e.g., using input information). In some examples, each of the weak decision voltage ranges includes an area around a hard decision threshold where voltage information of different values ​​of the channel output is likely to overlap.

[0113] According to some embodiments, the error correction component 815 performs an ECC operation and determines that the ECC operation failed.The memory channel output generated by a threshold expert in the plurality of threshold experts 810 can be generated based on the determination.

[0114] According to some embodiments, the memory device 220 includes a plurality of memory cells arranged in a plurality of word lines and a plurality of bit lines. According to some embodiments, the memory controller 210 reads a set of memory cells including the memory cells. In some examples, the memory controller 210 generates a codeword based on the reading. In some examples, the memory controller 210 determines that the ECC operation fails. In some examples, the memory controller 210 generates an updated codeword based on the improved memory channel output. In some examples, the memory controller 210 receives a read command from the host device 100. In some examples, the memory controller 210 reads voltage information from the memory cell in response to the read command, wherein the read information includes voltage information. In some examples, the memory controller 210 sends a response to the host device 100 based on the codeword.

[0115] In some examples, the memory controller 210 includes a processor. According to some embodiments, the memory controller 210 is configured to receive read information for a memory cell of the memory device 220. According to some embodiments, the memory controller 210 is configured to select a threshold expert from a plurality of threshold experts 810 based on the read information, wherein each threshold expert 810 corresponds to a different weak decision voltage range in a plurality of weak decision voltage ranges (e.g., as described herein, for example, with reference to Fig.11 1. In some embodiments, the memory controller 210 is configured to generate an estimate of the channel output based on the read information using the selected threshold expert. As described herein, in some examples, the host device 100 can be configured to receive a response from the memory controller 210 based at least in part on the estimate of the channel output.

[0116] Fig. 9 An example of an equalizer architecture using multiple threshold expert machine learning models according to aspects of the present disclosure is shown. The example shown includes a selection component 900, multiple threshold experts 905, read information 925, symbol estimation 955, meta network 960, and word line features 965.

[0117] Multiple threshold experts 905 may be included as referenced Figure 2 The plurality of threshold experts 905 are referenced Figure 8 and Fig.10 In one embodiment, the plurality of threshold experts 905 include a first threshold expert 910, a second threshold expert 915, and a third threshold expert 920. The first threshold expert 910, the second threshold expert 915, and the third threshold expert 920 are references to Fig.11 Examples of corresponding elements are described or include aspects thereof.

[0118] Select component 900 is reference Figure 8 Examples of corresponding elements described herein or including aspects thereof. Meta-network 960 is a reference Figure 8 Examples of corresponding elements described herein or including aspects thereof. Reading information 925 is a reference to Figure 6 Examples of corresponding elements described herein or including aspects thereof. Symbol estimation 955 is a reference Figure 6 Examples of corresponding elements described herein or including aspects thereof. Word line feature 965 is referenced Fig.10 Examples of corresponding elements are described or include aspects thereof.

[0119] exist Fig. 9 In an example of , the equalizer architecture includes a plurality of neural networks 905, each of which outputs soft symbol estimates 955 (e.g., to an error correction component) based on input read information 925 from a memory device (e.g., for the error correction component to perform an ECC operation on the output soft symbol estimates). The read information 925 may include cell voltage information 930, same-column neighbor voltage information 935, same-wordline neighbor voltage information 940, wordline HD threshold (e.g., all HD thresholds) information 945, and wordline index information 950, among others.

[0120] exist Fig. 9 In the example of , a sample (e.g., a read information sample, a detected voltage information sample, etc.) can be selected by a specific threshold expert in a plurality of threshold experts 905, wherein the specific threshold expert for each sample is selected by the selection component 900 based on the sample (e.g., based on the weak decision range in which the sample is located, as described in this article, for example, with reference to Fig.11 further described).

[0121] As described herein, each of the plurality of threshold experts 905 may include a shallow machine learning model. For example, each of the first threshold expert 910, the second threshold expert 915, the third threshold expert 920, etc. may include a shallow threshold expert machine learning model that is trained and configured to perform a classification task between two voltage states corresponding to a specific weak decision range. That is, the shallow threshold experts described herein may perform a binary classification operation for a specific weak decision range (e.g., as described herein, for example, with reference to Fig.11 600). Such a classification operation may be less complex relative to a classification operation for several possible voltage states (e.g., as performed by a more complex machine learning model (such as, for example, neural network 600). Therefore, due to reduced latency and memory parameters, various NAND architectures may support such a balanced architecture that implements multiple threshold experts 905 (e.g., multiple shallow threshold expert machine learning models that each perform a respective classification operation between two corresponding voltage states).

[0122] In some cases, Fig. 9 The example architecture of can be referred to as a multi-threshold architecture. Such an architecture can include multiple threshold expert models (e.g., such as a first threshold expert 910, a second threshold expert 915, a third threshold expert 920, etc.). Each threshold expert model can be assigned according to the HD threshold (e.g., as described in the referenced example herein). Fig.11 15. For example, the multi-threshold architecture may include 15 threshold expert models, but the present disclosure is not limited thereto. Each sample is received (e.g., analyzed, classified, etc.) by one of the threshold expert models in the plurality of threshold experts 905 depending on the detected sample voltage. The equalizer processes weak samples (e.g., samples within the weak decision range that are not easily classified into a voltage state by the equalizer).

[0123] The meta-network 960 can determine or generate parameters for each threshold expert model for classifying relevant samples received from the selection component 900 based on the corresponding HD proximity. In some examples, the meta-network 960 can identify word line features 965, and the meta-network 960 can be trained to generate parameters for each threshold expert model based on the word line features 965. The word line features 965 for the meta-network 960 can be based on the aggregate properties of sectors (e.g., in a memory), and can include characteristics or properties of one or more word lines in the memory. In some examples, the word line features 965 can be based on standard device operations, and the device can identify the word line features 965 without additional read operations.

[0124] Once the parameters are assigned to multiple threshold experts 905 by the meta-network 960, multiple threshold experts 905 can be used to perform error correction and denoising. The device can read the damaged codeword, read additional information describing the possibility of errors in a given unit, and estimate LLRs for each unit. The ECC decoder can then attempt to decode the codeword. The ECC decoder can search for the most likely codeword given the estimated LLRs. If the decoding is successful, the ECC can output the information word. However, if the decoding fails, the device can read additional information corresponding to the read information 925. The denoising algorithm can then perform error correction (e.g., for weak units) to change the value of the LLR, and the ECC decoder can attempt to decode the modified LLRs.

[0125] As used herein, a codeword may refer to information written to a device and may be part of an ECC codebook. As used herein, a corrupted codeword may refer to a word of similar size to a codeword but with some bits flipped (e.g., so the corrupted codeword may not be in the ECC codebook). As used herein, an LLR is generated for each cell and may refer to the natural logarithm of the ratio between the probability that a true bit value is 0 and the probability that a true bit value is 1 (e.g., ).

[0126] Fig.10 An example of a meta-network architecture according to aspects of the present disclosure is shown. A meta-network (or meta-model) in machine learning is a model that generates parameters for another model. In some examples, the meta-network can be implemented in firmware (e.g., based on C language code). Fig.10 In the example of , the meta-network 1000 can generate parameters for the MTEM model. Unlike a threshold expert that can be run (e.g., evaluated) for each cell (e.g., for each weak cell), the meta-network 1000 can be run (e.g., evaluated) once for each sector in the memory (e.g., where there can be about 34,000 cells in a sector). Therefore, a larger model can be used for the meta-network 1000 (e.g., a model with more parameters) with a minimal increase in complexity cost.

[0127] Because memory can be read in pages (e.g., a single bit out of N bits in a cell), meta-network 1000 can be designed to be efficient for single page reads. The input features for meta-network 1000 can be referred to as wordline features 1005 and can be obtained based on standard device operations (e.g., rather than using additional read operations). In some examples, wordline features 1005 may not be shared by the entire page. Once wordline features 1005 are obtained, meta-network 1000 can use a neural network to generate parameters for a threshold expert. In addition, once meta-network 1000 has generated (e.g., calculated) parameters for the threshold expert, a device (e.g., a MTEM equalizer) can be used as described above. Fig. 9 As described in, reading from memory is performed using a threshold expert for error correction.

[0128] In some implementations, the meta-network 1000 may include multiple neural networks 1010. Separate neural networks may be used to generate parameters for thresholds and corresponding threshold experts (e.g., corresponding linear threshold models). For example, one neural model may be used to generate parameters for threshold 0, while another neural model may be used to generate parameters for threshold 1. The input to the meta-network 1000 (e.g., word line features 1005) may be shared among all neural networks 1010. In addition, the output of the neural network may be parameters (e.g., 5 parameters) for corresponding thresholds and threshold experts (e.g., used to set or configure the threshold experts).

[0129] In some examples, the parameters generated for the threshold expert can include weights of upper cell voltage features (e.g., nearest neighbor cells with higher wordline indexes along a pillar), weights of target cell voltage features, weights of lower cell voltage features (e.g., nearest neighbor cells with lower wordline indexes along a pillar), weights of threshold voltage features, and bias term values. The architecture of a neural network (e.g., neural network 1010) for generating parameters for the threshold expert can include a fully connected feedforward neural network, two hidden layers (e.g., one hidden layer has 32 neurons and the other hidden layer has 8 neurons), rectified linear unit (RELU) activations after each hidden layer, and a fully linear layer as an output layer.

[0130] Once the parameters are assigned to the threshold expert, the threshold expert can be used for error correction. The threshold expert can accept one or more inputs, including the target cell voltage, the neighbor voltage along the upper column, the neighbor voltage along the lower column, and the threshold reference voltage (e.g., the voltage used for hard decisions without an equalizer). Based on these inputs, the threshold expert can generate a single output F w (X) = b + ∑ i w i ·X i , where b is the bias term, and wi is the weight assigned to input feature i. In some examples, the threshold expert can be trained using standard machine learning techniques so that F w (X) Estimate LLR.

[0131] Fig.11 An example of a sample denoising scheme according to aspects of the present disclosure is shown. The example shown includes a hierarchy 1100, a threshold 1105, a strong decision voltage range 1110, a weak decision voltage range 1115, and a plurality of threshold experts 1120. The plurality of threshold experts 1120 are reference Figure 8 and Fig. 9 In one embodiment, the plurality of threshold experts 1120 include a first threshold expert 1125, a second threshold expert 1130, and a third threshold expert 1135. The plurality of threshold experts 1120 may be included in the example of reference Figure 2 The first threshold expert 1125 is a reference Fig. 9 The second threshold expert 1130 is a reference to the corresponding element of the example or includes aspects thereof. Fig. 9 The third threshold expert 1135 is a reference to the corresponding element of the example or includes its aspects. Fig. 9 Examples of corresponding elements are described or include aspects thereof.

[0132] To read information from a memory device, the voltage of each cell is measured and the voltage level stored in the cell is inferred. The bits can then be recovered. There may be a trade-off between the number of voltage levels and memory reliability. The greater the number of bits per cell, the more information can be stored on the device and the more bits in each cell. Alternatively, voltages representing different levels can be packed more tightly together because a greater number of distinguishable voltages are used within the same dynamic range. Therefore, noise in cell programming or cell reading is more likely to change the voltage of one level to another voltage representing a different level, thereby manifesting an error when reading the cell.

[0133] There are multiple sources of noise in memory devices that can lead to erroneous reading of information, such as write noise, disturb noise, aging, and read operations. Write noise is attributed to the programming process and is found when the voltage of a cell is different from the expected voltage immediately after programming. Interference noise is attributed to the voltage of a cell changing due to programming different neighboring cells. Programming a cell causes disturbances that affect other cells. Aging means that the more times a device is written and read, the more noise it will have. In addition, the more time between programming of a cell, the more noise the cell will generate. Similarly, the operation of reading a cell can cause noise and disturbances.

[0134] exist Fig.11In the example of , a sample (e.g., a read information sample, a detected voltage information sample, etc.) can be selected by a threshold expert (e.g., a first threshold expert 1125, a second threshold expert 1130, a third threshold expert 1135, etc.) selected by a selection component from a plurality of threshold experts 1120. As described herein, a particular threshold expert for each sample is selected (e.g., by a selection component) based on a characteristic of the sample (e.g., a detected value, such as a detected voltage). For example, the threshold expert is selected based on a weak decision range 1115 in which the sample is located.

[0135] For example, as described herein, a memory device (e.g., a NAND flash memory device) may allow several bits of data to be stored in each memory cell. A memory cell in which multiple bits of data are stored may be referred to as a multi-layer memory cell. A multi-layer memory cell divides the threshold voltage range of the memory cell into several voltage states, and the data value written to the memory cell is extracted using a memory cell voltage level (e.g., such as "level 0," "level 1," ... "level 7").

[0136] To read information from a memory device, the voltage of each cell is measured and the voltage level stored in the cell (e.g., such as "level 0", "level 1", ... "level 7") is inferred. The bits can then be recovered. The greater the number of bits per cell, the more information that can be stored on the device. However, the voltages representing different levels for such cells (e.g., voltages representing "level 0", "level 1", ... "level 7") can be packed more tightly together because a greater number of distinguishable voltages are used within the same dynamic range. As described herein, when information is read from a cell (e.g., a NAND cell), the read information may be corrupted by noise (e.g., depending on the characteristics of the medium), which may result in an ambiguous value for the read information (e.g., the read voltage level may be within a weak decision voltage range 1115 between two voltage values).

[0137] exist Fig.11In the example of , eight levels 1100 (e.g., “Level 0”, “Level 1”, … “Level 7”) are shown, which can be represented by three bits (e.g., “000” = “Level 0”; “001” = “Level 1”; … “111” = “Level 7”). The HD threshold 1105 may exist between two adjacent levels 1100. In addition, within the dynamic range spanning “Level 0” to “Level 7”, a strong decision range 1110 and a weak decision range 1115 may exist. The strong decision range 1110 may refer to a range (e.g., a voltage range) in which the level of a sample is easily determined (e.g., by a voltage detector). The weak decision range 1115 may include a range around the HD threshold 1105 in which the level of a sample is not easily determined (e.g., by a voltage detector). For example, the weak decision range 1115 may include values ​​(e.g., voltage values) immediately before and after the HD threshold 1105.

[0138] In general, the levels 1100 may be distinguishable based on the HD threshold 1105, and the levels 1100 may include a portion of a strong decision range 1110 and a weak decision range 1115. Read information (e.g., detected voltage) within the strong decision range 1110 may be easily classified into the corresponding level 1100 (e.g., by a voltage detector without using a threshold expert). Read information (e.g., detected voltage) within the weak decision range 1115 may be passed to a selection component, where the selection component may select a threshold expert to classify the read information into one of two adjacent levels 1100. Fig.11 In the example of , HD threshold 1105-b can be located between level 1100-a and level 1100-b. Weak decision range 1115 can include a range between level 1100-a and level 1100-b, where further analysis of the reading information is performed via a corresponding threshold expert in multiple threshold experts 1120. As described herein, weak decision range 1115 can include a range where reading information of different values ​​output by a channel is likely to overlap (e.g., so that a threshold expert can be used to classify the reading information as one of two associated adjacent values ​​to obtain improved accuracy and reliability).

[0139] Thus, strong samples or samples within the strong decision range 1110 may have a channel output estimated based on the voltage detector (e.g., without using a threshold expert because the sample is within the strong decision range 1110, where the accuracy of the estimated channel output is relatively high). However, weak samples or samples within the weak decision range 1115 may have a channel output estimated by a threshold expert selected by the selection component (e.g., where the threshold expert is selected based on an HD threshold classification operation associated with the weak decision range 1115 where the detected read information is located).

[0140] Furthermore, as described herein, such an equalization operation may be performed after an ECC operation fails to generate an updated codeword for a subsequent ECC operation attempt. In the event that an ECC operation succeeds, an equalization operation does not necessarily need to be performed.

[0141] In the example scenario, the first threshold expert 1125 is trained based on samples near HD threshold 0, the second threshold expert 1130 is trained based on samples near HD threshold 1, and so on (e.g., up to the third threshold expert 1135 trained based on samples near HD threshold 6). Each threshold expert model solves a binary classification task. During inference, each sample passes through one of the threshold expert models. As an example, the pseudo code for the single threshold mode is described below.

[0142] Single threshold mode pseudocode:

[0143]

[0144]

[0145] The neural network nature of the threshold models is just an example, and any other type of machine learning model (decision tree, random forest, etc.) can be used. Embodiments of the present disclosure are based on multiple shallow threshold expert machine learning models. The multi-threshold architecture includes several machine learning models, one machine learning model for each HD threshold (for example, or in some cases, as described herein, one threshold expert for each subset of all HD thresholds across a given dynamic range). Each threshold expert is a shallow machine learning model, and each threshold expert can solve a binary classification task. During inference, each sample passes through one of the threshold models. The threshold expert is selected by a selection component (e.g., a threshold model selection module) based on the voltage of the memory cell and / or the closest HD threshold.

[0146] The multi-threshold architecture of the present disclosure produces a set of shallow machine learning models that can operate with low memory usage (memory usage is less than ECC operation memory usage) and low latency. The balanced model can be implemented in a firmware version. Therefore, the balanced model can be used by any NAND flash memory architecture.

[0147] The present disclosure reduces latency and improves accuracy, wherein each threshold expert model is shallow and includes binary outputs. Additionally or alternatively, relevant samples are used for each training model, thereby improving accuracy. For example, Table 1 shows example checks of TLC and quad-layer cell (QLC) data sets. Table 1 summarizes the bit error rate (BER) reduction achieved for each data set.

[0148] Dataset Training set Test Set BER reduction [%] TLC 50 blocks 22 blocks 54.08% QLC 73 blocks 31 blocks 17.8%

[0149] Table 1

[0150] Fig.11 A selection component (e.g., a threshold model selection module, such as reference Figure 8 and Fig. 9 1 and 12. The present invention also provides aspects of the selection component 805 and the selection component 900 described above. For example, in the selection component, a threshold expert is selected for each sample within a weak decision range 1115 (e.g., within a range 1115 between two states or between two "levels"). A threshold expert can be selected that classifies between two states or levels (e.g., between two voltage states). In some cases, the weak decision range can be referred to as a range around an HD threshold (e.g., the weak decision range 1115 can include and surround one of "threshold 0", "threshold 1", ... "threshold 6").

[0151] In an example scenario, the voltage of each sample is determined (e.g., via the voltage detector 800). Then, for samples having a voltage within a weak decision range 1115, the HD is determined based on the voltage closest to the determined sample voltage (e.g., the closest voltage state). The characteristics of the sample are then provided to a corresponding threshold module (e.g., a threshold expert trained or configured to perform a classification operation on the weak decision range 1115 associated with the voltage of the sample determined by the voltage detector). Improved parameters for each sample can then be determined because each sample can be classified by a threshold expert that is trained and configured to classify samples within a specific weak decision range 1115 that includes the detected sample voltage. In addition, due to the reduced latency and memory parameters associated with multiple shallow threshold experts 1120, various NAND architectures can support such a balanced architecture that implements multiple threshold experts 1120.

[0152] also, Fig.11 The examples are provided for illustrative purposes and are not intended to limit the scope of the present disclosure. Fig.11An example is shown in which parameters such as level estimates may be determined, however other estimates (e.g., such as likelihood bits, soft symbols, etc.) may also be used. In another example embodiment of the present disclosure, the number of threshold experts may be reduced (e.g., because the threshold expert may perform classification operations on two weak decision ranges 1115). For example, samples within an HD threshold behave similarly. Therefore, the same threshold model may be used for samples in different HD (e.g., weak decision ranges 1115) neighbors, thereby reducing the number of threshold models in the equalization module. For example, a first threshold expert 1125 may be used for classification operations associated with both a weak decision range 1115 between "level 0" and "level 1" and a weak decision range 1115 between "level 1" and "level 2". In such an example, a second threshold expert 1130 may be used for classification operations associated with both a weak decision range 1115 between "level 2" and "level 3" and a weak decision range 1115 between "level 3" and "level 4".

[0153] Each threshold expert can be a shallow neural network (e.g., where each shallow neural network can be trained or configured for classification operations for different HD thresholds or different weak decision ranges 1115). Some embodiments of the present disclosure use a logistic regression model, but other architectures can also be used. An accuracy-model simplicity tradeoff can be obtained.

[0154] Thus, embodiments of the present disclosure can reduce the total number of threshold experts and still reduce latency and memory usage (e.g., via the use of multiple shallow threshold expert networks that are trained or configured for one or more specific classification operations, rather than the use of a single complex machine learning network that performs classification operations for a wide range of levels or states).

[0155] As described herein, the present disclosure includes the following embodiments.

[0156] A method for flash memory denoising using a machine learning model is described. An embodiment of the method is configured to: receive first information for a plurality of memory cells of a memory device and second information for a memory cell in the plurality of memory cells, generate one or more threshold expert parameters for a threshold expert based on the first information, select a threshold expert from the plurality of threshold experts based on the second information, wherein each of the threshold experts corresponds to at least one of a plurality of weak decision voltage ranges for the memory cell, and generate a memory channel output based on the second information using the selected threshold expert and the one or more threshold expert parameters.

[0157] Some examples of the above method further include determining that a voltage value of the second information is within at least one weak decision voltage range corresponding to a threshold expert, wherein the threshold expert is selected based on the determination.

[0158] In some examples, each of the weak decision voltage ranges includes a region between two hard decision thresholds where voltage information of different values ​​output by a channel is likely to overlap.

[0159] In some examples, each of the threshold experts is trained to operate using input information from a corresponding weak decision voltage range of the plurality of weak decision voltage ranges.

[0160] Some examples of the above method also include performing an ECC operation and determining that the ECC operation failed, wherein the memory channel output is generated based on the determination.

[0161] In some examples, the plurality of memory cells comprises a word line or a subset of word lines of a memory device.

[0162] In some examples, the first information includes a number of PE cycles, a layer of a target page, a layer index, a very strong error ratio, a normal strong error ratio, a normal weak error ratio, a very weak error ratio, one or more threshold voltages, or any combination thereof.

[0163] In some examples, the second information includes an upper cell voltage, a target cell voltage, a lower cell voltage, a threshold voltage, or any combination thereof.

[0164] In some examples, the one or more threshold expert parameters include a weight of an upper cell voltage characteristic, a weight of a target cell voltage characteristic, a weight of a lower cell voltage characteristic, a weight of a threshold voltage characteristic, a bias term value, or any combination thereof.

[0165] Some examples of the above method also include generating a threshold expert parameter for each of the plurality of threshold experts.

[0166] An apparatus for flash memory denoising using a machine learning model is described. The apparatus includes: a memory device including a plurality of memory cells arranged in a plurality of word lines; a meta-network configured to generate one or more threshold expert parameters for a threshold expert based on first information for a word line in the plurality of word lines; and a plurality of threshold experts configured to generate a memory channel output based on second information from the memory cells in the word line and the one or more threshold expert parameters.

[0167] In some examples, the meta-network includes multiple models corresponding to multiple threshold experts, respectively.

[0168] Some examples of the above apparatus further include a threshold expert selection component configured to select a threshold expert from a plurality of threshold experts based on the second information.

[0169] In some examples, each of the multiple threshold experts includes a linear machine learning model.

[0170] In some examples, the meta-network includes a fully connected feed-forward neural network.

[0171] Some examples of the above apparatus also include a decoder configured to perform an ECC operation, wherein the memory channel output is generated based on a result of the ECC operation.

[0172] A non-transitory computer-readable medium for flash memory denoising using a machine learning model is described. In some examples, the code includes instructions executable by a processor to: receive first information for a plurality of memory cells of a memory device and second information for a memory cell in the plurality of memory cells, generate one or more threshold expert parameters for a threshold expert based on the first information, select a threshold expert from the plurality of threshold experts based on the second information, wherein each of the threshold experts corresponds to at least one of a plurality of weak decision voltage ranges for the memory cell, and generate a memory channel output based on the second information using the selected threshold expert and the one or more threshold expert parameters.

[0173] In some examples, the code includes instructions executable by the processor to determine that a voltage value of the second information is within at least one weak decision voltage range corresponding to a threshold expert, wherein the threshold expert is selected based on the determination.

[0174] In some examples, the code includes instructions executable by a processor to perform an ECC operation and determine that the ECC operation failed, wherein a memory channel output is generated based on the determination.

[0175] In some examples, the code includes instructions executable by a processor to generate threshold expert parameters for each of a plurality of threshold experts.

[0176] The descriptions and drawings given herein represent example configurations and do not represent all implementations within the scope of the claims. For example, operations and steps may be rearranged, combined, or otherwise modified. In addition, structures and devices may be represented in the form of block diagrams to represent the relationships between components and avoid blurring the concepts described. Similar components or features may have the same name, but may have different reference numerals corresponding to different drawings.

[0177] Some modifications to the present disclosure may be apparent to those skilled in the art, and the principles defined herein may be applied to other variations without departing from the scope of the present disclosure. Therefore, the present disclosure is not limited to the examples and designs described herein, but will conform to the broadest scope consistent with the principles and novel features disclosed herein.

[0178] The described systems and methods may be implemented or performed by a device including a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof. A general purpose processor may be a microprocessor, a conventional processor, a controller, a microcontroller, or a state machine. The processor may also be implemented as a combination of computing devices (e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a DSP core, or any other such configuration). Therefore, the functions described herein may be implemented in hardware or software and may be performed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored in a computer readable medium in the form of instructions or codes.

[0179] Computer-readable media include both non-transitory computer storage media and communication media, and communication media include any media that facilitates the transmission of code or data. Non-transitory storage media can be any available media that can be accessed by a computer. For example, non-transitory computer-readable media can include random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact disk (CD) or other optical disk storage, magnetic disk storage, or any other non-transitory medium for carrying or storing data or code.

[0180] In addition, the connecting component may be appropriately referred to as a computer-readable medium. For example, if the code or data is sent from a website, server or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technology (such as infrared, radio or microwave signals), the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technology is included in the definition of medium. Combinations of media are also included within the scope of computer-readable media.

[0181] In the present disclosure and the appended claims, the word "or" indicates an inclusive list, such that a list such as X, Y, or Z means X or Y or Z or XY or XZ or YZ or XYZ. Furthermore, the phrase "based on" is not used to represent a closed set of conditions. For example, a step described as "based on condition A" may be based on both condition A and condition B. In other words, the phrase "based on" should be interpreted to mean "based at least in part on". Furthermore, the words "a" or "an" indicate "at least one".

Claims

1. A method for flash memory denoising, include: receiving first information for a plurality of memory cells of a memory device and second information for a memory cell of the plurality of memory cells; generating one or more threshold expert parameters for the threshold expert based on the first information; selecting a threshold expert from a plurality of threshold experts based on second information, wherein each of the plurality of threshold experts corresponds to at least one weak decision voltage range of a plurality of weak decision voltage ranges of the memory cell; as well as A memory channel output is generated based on the second information using the selected threshold expert and the one or more threshold expert parameters.

2. The method according to claim 1, further comprising: include: It is determined that a voltage value of the second information is within at least one weak decision voltage range corresponding to the threshold expert, wherein the threshold expert is selected based on the determination.

3. The method according to claim 1, in: Each of the plurality of weak decision voltage ranges includes a region between two hard decision thresholds where voltage information of different values ​​output by the channels is likely to overlap.

4. The method according to claim 1, in: Each of the plurality of threshold experts is trained to operate using input information from a corresponding weak decision voltage range of the plurality of weak decision voltage ranges.

5. The method according to claim 1, further comprising: include: Perform error correction code ECC operations; A determination is made that the ECC operation has failed, wherein the memory channel output is generated based on the determination.

6. The method according to claim 1, in: The plurality of memory cells comprises a word line or a subset of word lines of the memory device.

7. The method according to claim 1, in: The first information includes a number of program-erase cycles, a layer of a target page, a layer index, a very strong error ratio, a normal strong error ratio, a normal weak error ratio, a very weak error ratio, one or more threshold voltages, or any combination thereof.

8. The method according to claim 1, in: The second information includes an upper cell voltage, a target cell voltage, a lower cell voltage, a threshold voltage, or any combination thereof.

9. The method according to claim 1, in: The one or more threshold expert parameters include a weight of an upper cell voltage characteristic, a weight of a target cell voltage characteristic, a weight of a lower cell voltage characteristic, a weight of a threshold voltage characteristic, a bias term value, or any combination thereof.

10. The method according to claim 1, further comprising: include: Threshold expert parameters are generated for each of the plurality of threshold experts.

11. A device, include: A memory device comprising a plurality of memory cells arranged in a plurality of word lines; a meta-network configured to generate one or more threshold expert parameters for a threshold expert based on first information for a wordline of the plurality of wordlines; as well as A plurality of threshold experts are configured to generate memory channel outputs based on second information from memory cells in the word line and the one or more threshold expert parameters.

12. The device according to claim 11, in, The meta-network includes a plurality of models corresponding to the plurality of threshold experts respectively.

13. The device according to claim 11, further comprising: include: The threshold expert selection component is configured to select the threshold expert from the plurality of threshold experts based on second information.

14. The device according to claim 11, in, Each of the plurality of threshold experts comprises a linear machine learning model.

15. The device according to claim 11, in, The meta-network includes a fully connected feed-forward neural network.

16. The device according to claim 11, further comprising: include: The decoder is configured to perform an error correction code (ECC) operation, wherein the memory channel output is generated based on a result of the ECC operation.

17. A non-transitory computer-readable medium storing executable instructions, which, when executed by a processing device, cause the processing device to perform operations comprising: receiving first information for a plurality of memory cells of a memory device and second information for a memory cell of the plurality of memory cells; generating one or more threshold expert parameters for the threshold expert based on the first information; selecting a threshold expert from a plurality of threshold experts based on second information, wherein each of the plurality of threshold experts corresponds to at least one weak decision voltage range of a plurality of weak decision voltage ranges of the memory cell; as well as A memory channel output is generated based on the second information using the selected threshold expert and the one or more threshold expert parameters.

18. The non-transitory computer readable medium of claim 17, in, The instructions are further executable to cause the processing device to perform operations including: It is determined that a voltage value of the second information is within at least one weak decision voltage range corresponding to the threshold expert, wherein the threshold expert is selected based on the determination.

19. The non-transitory computer readable medium of claim 17, in, The instructions are further executable to cause the processing device to perform operations including: performing error correction code (ECC) operations; and A determination is made that the ECC operation has failed, wherein the memory channel output is generated based on the determination.

20. The non-transitory computer readable medium of claim 17, in, The instructions are further executable to cause the processing device to perform operations including: Threshold expert parameters are generated for each of the plurality of threshold experts.