Memory device for training neural networks

By training neural networks in a memory bank within a volatile memory device, the problem of requiring large amounts of computing resources for training neural networks in the prior art is solved, achieving efficient and automated neural network training, and reducing dependence on host systems and power consumption.

CN115461758BActive Publication Date: 2026-11-17MICRON TECHNOLOGY INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202180031503.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2020-05-14
Filing Date
2021-04-26
Publication Date
2026-11-17
Estimated Expiration
2041-04-26

AI Technical Summary

Technical Problem

In the existing technology, training neural networks requires a large amount of computing resources and dedicated circuit systems, making it difficult to efficiently manage and train neural network data within a memory device.

Method used

Data management and training of neural networks can be implemented in volatile memory devices such as DRAM. Different neural networks can be trained through operations within the memory bank, and the storage and training operations of neural networks can be controlled by a control circuit system, avoiding dependence on external hosts and dedicated hardware.

Benefits of technology

This enables training neural networks within a memory device without requiring significant computational resources, reducing dependence on the host system and power consumption, and improving training efficiency and automation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115461758B_ABST
    Figure CN115461758B_ABST
Patent Text Reader

Abstract

Methods, systems, and apparatuses related to training neural networks are described. For example, data management and training of one or more neural networks can be implemented within a memory device, such as a dynamic random access memory (DRAM) device. Therefore, neural networks can be trained without dedicated circuitry and / or without significant computational resources. One or more neural networks can be written to or stored in memory banks of the memory device, and operations for training different neural networks located in different banks of the memory device can be performed within or adjacent to those memory banks. This data management and training can be performed within the memory system without involving host devices, processors, or accelerators outside the memory system. The trained network can then be read from the memory system and used for inference or other operations on external devices.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure generally relates to semiconductor memories and methods, and more specifically, to memory devices for training neural networks. Background Technology

[0002] Memory devices are typically provided internally in computers or other electronic systems, in semiconductors, or in integrated circuits. Many different types of memory exist, including volatile and non-volatile memory. Volatile memory may require power to maintain its data (e.g., host data, error data, etc.) and includes Random Access Memory (RAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Synchronous Dynamic Random Access Memory (SDRAM), and Thyristor Random Access Memory (TRAM), among others. Non-volatile memory provides permanent data by retaining stored data when no power is supplied and includes NAND flash memory, NOR flash memory, and resistive variable memory such as Phase-Change Random Access Memory (PCRAM), Resistive Random Access Memory (RRAM), and Magnetoresistive Random Access Memory (MRAM), such as Spin Torque Transfer Random Access Memory (STT RAM), among others.

[0003] A memory device may be coupled to a host computer (e.g., a host computing device) to store data, commands, and / or instructions for use by the host computer or electronic system during operation. For example, data, commands, and / or instructions may be transferred between the host computer and the memory device during operation of the computing or other electronic system. Attached Figure Description

[0004] Figure 1 This is a functional block diagram of a device comprising a host and a memory device according to several embodiments of the present disclosure.

[0005] Figure 2A It is a functional block diagram in the form of a memory device that includes a control circuit system and multiple memory banks for storing neural networks.

[0006] Figure 2B It is another functional block diagram in the form of a memory device that includes a control circuit system and multiple memory banks storing multiple neural networks.

[0007] Figure 3 It is another functional block diagram in the form of a memory device that includes a control circuit system and multiple memory banks for storing neural networks.

[0008] Figure 4 This is a flowchart illustrating an example method of a memory device corresponding to a training neural network according to several embodiments of the present disclosure.

[0009] Figure 5 This is a flowchart illustrating another example method of a memory device corresponding to a training neural network according to several embodiments of the present disclosure.

[0010] Figure 6 This is a schematic diagram illustrating a portion of a memory array including a sensing circuit system according to several embodiments of the present disclosure. Detailed Implementation

[0011] Methods, systems, and apparatuses relating to training neural networks are described. For example, data management and training of one or more neural networks can be implemented within a memory device (e.g., a dynamic random access memory (DRAM) device). Therefore, neural networks can be trained without dedicated circuitry and / or without significant computational resources. One or more neural networks can be written to or stored in memory banks of the memory device, and operations for training different neural networks located in different groups within the memory device can be performed within or adjacent to those memory banks. This data management and training can occur within the memory system without involving host devices, processors, or accelerators outside the memory system. The trained network can then be read from the memory system and used for inference or other operations on external devices.

[0012] Neural networks can contain a set of instructions that can be executed to identify patterns in data. Some neural networks can be used to identify lower-level relationships in a dataset in a way that mimics the operation of the human brain. Neural networks can adapt to changing or altered inputs, allowing them to produce the best possible results without redesigning the output criteria.

[0013] A neural network can consist of multiple neurons, which can be represented by one or more equations. In the context of a neural network, a neuron can receive a large number of numbers or vectors as input and produce an output based on the properties of the neural network. For example, a neuron can receive X... k There are *k* inputs, where *k* corresponds to the index of the input. For each input, the neuron can assign a weight vector W to the input. k In some embodiments, the weight vector can make a neuron in the neural network different from one or more different neurons in the network. In some neural networks, the corresponding input vector can be multiplied by the corresponding weight vector to obtain a value, as shown by Equation 1, which illustrates an example of a linear combination of the input vector and the weight vector.

[0014]

[0015] Equation 1

[0016] In some neural networks, a nonlinear function (e.g., an activation function) can be applied to the value f(x1, x2) produced by Equation 1. An example of a nonlinear function that can be applied to the value produced by Equation 1 is the Modified Linear Unit (ReLU) function. Applying the ReLU function, as shown in Equation 2, produces a value input to the function if the value is greater than zero, or zero if the value input to the function is less than zero. The ReLU function is used here only as an illustrative example of an activation function and is not intended to be restrictive. Other non-restrictive examples of activation functions that can be applied in the context of neural networks include sigmoid functions, binary step functions, linear activation functions, hyperbolic functions, leaky ReLU functions, parametric ReLU functions, softmax functions, and / or swish functions, etc.

[0017]

[0018] Equation 2

[0019] During the training of a neural network, the input vector and / or weight vector can be changed to "tune" the network. In one instance, the neural network can be initialized with random weights. Over time, the weights can be adjusted to improve the accuracy of the neural network. This results in a neural network with high accuracy over time.

[0020] Neural networks have a wide range of applications. For example, they can be used for system recognition and control (vehicle control, trajectory prediction, process control, natural resource management), quantum chemistry, general game theory, pattern recognition (radar systems, face recognition, signal classification, 3D reconstruction, object recognition, and more), sequence recognition (gesture, speech, handwritten and printed text recognition), medical diagnosis, finance (e.g., automated trading systems), data mining, visualization, machine translation, social network filtering and / or email spam filtering, and much more.

[0021] Due to the computational resources required for some neural networks, in some methods, the neural network is deployed in a computing system (e.g., a mainframe computing system (e.g., a desktop computer, a supercomputer, etc.) or a cloud computing environment). In such methods, as part of the operation of training the neural network, the data processed by the neural network can be stored in memory resources (e.g., NAND flash memory devices), and processing resources (e.g., a central processing unit) can access the data and use the neural network to execute instructions for processing the data. As part of the neural network training, some methods may also utilize dedicated hardware, such as field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs).

[0022] In contrast, the embodiments described herein are geared towards data management and training of one or more neural networks within a volatile memory device (e.g., a dynamic random access memory (DRAM) device). Therefore, the embodiments described herein can allow training of neural networks without dedicated circuitry and / or without significant computational resources. As described in more detail herein, embodiments of this disclosure include writing one or more neural networks into a memory bank of the memory device and performing operations to train different neural networks located in different groups within the memory device using the neural networks. For example, in some embodiments, a first neural network may be written into a first memory bank (or a first subset of the memory bank) and a second neural network may be written into a second memory bank (or a second subset of the memory bank). The first or second neural network can be used to train the other of the first or second neural networks. Furthermore, the embodiments described herein can allow training of the first or second neural network "on-chip" (e.g., without hindering a host coupled to the memory device and / or without transferring the neural network to a location outside the memory device).

[0023] In the following detailed description of this disclosure, reference is made to the accompanying drawings, which form a part of this disclosure, and the drawings illustrate by way of illustration one or more embodiments of this disclosure. These embodiments are described in sufficient detail to enable those skilled in the art to practice embodiments of this disclosure, and it should be understood that other embodiments may be utilized and that process, electrical, and structural changes may be made without departing from the scope of this disclosure.

[0024] As used herein, designators such as “X,” “N,” and “M,” which are specific references to the reference numerals in the drawings, indicate that a plurality of such specific features may be included. It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms “a / an” and “the” may include both singular and plural references. Additionally, “a plurality,” “at least one,” and “one or more” (e.g., a plurality of memory banks) may refer to one or more memory banks, while “a plurality” is intended to refer to more than one such thing.

[0025] Furthermore, the words “may” and “may” are used throughout this application in a permissive sense (i.e., possible, able) rather than in a mandatory sense (i.e., must). The term “comprising” and its derivatives mean “including but not limited to”. Depending on the context, the term “coupled / coupling” means a physical, direct or indirect connection or access to and movement (transmission) of commands and / or data. Depending on the context, the terms “data” and “data value” are used interchangeably and may have the same meaning herein.

[0026] The diagrams in this document follow a numbering rule, where the first one or more digits correspond to the diagram number, and the remaining digits identify elements or components within the diagram. Similar elements or components between different diagrams can be identified by using similar digits. For example, 104 could represent... Figure 1 The element “04” in Figure 2 is used, and a similar element may be represented as 204 in Figure 2. Generally, a single element symbol may be used herein to refer to multiple similar elements or components or groups of elements or components. For example, multiple reference elements 221-1 to 221-N (or, in an alternative, 221-1, ..., 221-N) may be collectively referred to as 221. As will be understood, elements shown in various embodiments herein may be added, interchanged, and / or removed to provide several additional embodiments of this disclosure. Furthermore, the scale and / or relative dimensions of the elements provided in the figures are intended to illustrate certain embodiments of this disclosure and should not be construed as limiting.

[0027] Figure 1 This is a functional block diagram of a computing system 100 comprising a device including a host 102 and a memory device 104, according to several embodiments of the present disclosure. As used herein, "device" may refer to, but is not limited to, any of a variety of structures or combinations thereof, such as a circuit or circuit system, one or more dies, one or more modules, one or more devices, or one or more systems. The memory device 104 may include one or more memory modules (e.g., a single in-line memory module, a dual in-line memory module, etc.). The memory device 104 may include volatile memory and / or non-volatile memory. In several embodiments, the memory device 104 may include a multi-chip device. A multi-chip device may include several different memory types and / or memory modules. For example, a memory system may include non-volatile or volatile memory on any type of module. Figure 1 As shown, the computing system 100 may include a control circuitry system 120 (which may include a logic circuitry system 122 and a memory resource 124), a memory array 130, and a sensing circuitry system 150 (e.g., a sensing (SENSE) 150). This document incorporates... Figure 6 Examples of the sensing circuitry system 150 are described in more detail below. For example, in several embodiments, the sensing circuitry system 150 may include several sensing amplifiers and corresponding computing components that can act as accumulators and be used to perform neural network training operations using trained and untrained neural networks stored in the memory array 130. Additionally, each of the components (e.g., host 102, control circuitry system 120, logic circuitry system 122, memory resource 124, memory array 130, and / or sensing circuitry system 150) may be individually referred to herein as a “device.” The control circuitry system 120 may be referred to herein as a “processing device” or a “processing unit.”

[0028] Memory device 104 may provide main memory for computing system 100 or may be used as additional memory or storage device throughout computing system 100. Memory device 104 may include one or more memory arrays 130 (e.g., memory cell arrays), which may contain volatile and / or non-volatile memory cells. For example, memory array 130 may be a flash array with a NAND architecture. Embodiments are not limited to a particular type of memory device. For example, memory device 104 may include RAM, ROM, DRAM, SDRAM, PCRAM, RRAM, and flash memory, etc.

[0029] In embodiments where memory device 104 includes non-volatile memory, memory device 104 may include flash memory devices, such as NAND or NOR flash memory devices. However, embodiments are not limited thereto, and memory device 104 may include other non-volatile memory devices, such as non-volatile random access memory devices (e.g., NVRAM, ReRAM, FeRAM, MRAM, PCM), "emerging" memory devices such as variable resistance (e.g., 3-D crosspoint (3D XP) memory devices), memory devices including self-selected memory (SSM) cell arrays, etc., or combinations thereof.

[0030] Variable resistance memory devices can combine stackable cross-grid data access arrays to perform bit storage based on changes in bulk resistance. Furthermore, compared to many flash-based memories, variable resistance non-volatile memories allow for in-situ write operations, where non-volatile memory cells can be programmed without pre-erasing them. Compared to flash-based memories and variable resistance memories, custom memory cells can comprise memory cells made of a single chalcogenide material, serving as both switches and storage elements for the memory cells.

[0031] like Figure 1 As described, host 102 may be coupled to memory device 104. In several embodiments, memory device 104 may be coupled to host 102 via one or more channels (e.g., channel 103). Figure 1 In this configuration, memory device 104 is coupled to host 102 via channel 103, and memory device 104 control circuitry system 120 is coupled to memory array 130 via channel 107. Host 102 may be a host system, such as a personal laptop, desktop computer, digital camera, smartphone, memory card reader and / or Internet of Things (IoT) enabled device, as well as various other types of host systems.

[0032] Host 102 may include a system motherboard and / or backplane and may include memory access devices, such as a processor (or processing device). Those skilled in the art will understand that a "processor" can be one or more processors, such as a parallel processing system, several coprocessors, etc. Computing system 100 may include a separate integrated circuit or both host 102 and memory device 104, and memory array 130 may be on the same integrated circuit. For example, computing system 100 may be a server system and / or a high-performance computing (HPC) system and / or a portion thereof. Although Figure 1 The examples shown illustrate systems with a von Neumann architecture, but embodiments of this disclosure can be implemented in non-von Neumann architectures that may not include one or more components typically associated with a von Neumann architecture (e.g., CPU, ALU, etc.).

[0033] The memory device 104, shown in more detail in Figure 2 herein, may include a control circuitry system 120, which may include a logic circuitry system 122 and memory resources 124. The logic circuitry system 122 may be provided herein in the form of an integrated circuit, such as an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a reduced instruction set computing device (RISC), an advanced RISC machine, a system-on-a-chip, or other combination of hardware and / or circuitry systems configured to perform the operations described in more detail. In some embodiments, the logic circuitry system 122 may include one or more processors (e.g., processing devices, processing units, etc.).

[0034] The logic circuit system 122 can perform operations to control access to and from the memory array 130 and / or the sensing circuit system 150. For example, as described herein in conjunction with Figures 2 and 3, the logic circuit system 122 can perform operations to control the storage of one or more neural networks within the memory array 130. In some embodiments, the logic circuit system 122 can receive a command from the host 102 and can control the storage of the neural network in the memory array 130 in response to receiving the command. However, embodiments are not limited thereto, and in some embodiments, the logic circuit system 122 can cause the neural network to be stored in the memory array 130 without a command from the host 102. As described in more detail herein in conjunction with Figures 2 and 3, at least one of the stored neural networks may be trained before being stored in the memory array 130. Similarly, in some embodiments, at least one of the stored neural networks may be untrained before being stored in the memory array 130.

[0035] Once the neural network is stored in the memory array 130, the logic circuit system 122 can control the initiation of operations using the stored neural network. For example, in some embodiments, the logic circuit system 122 can control the initiation of operations to train other neural networks (e.g., one or more untrained neural networks) stored in the memory array 130 using one or more stored neural networks (e.g., one or more trained neural networks). However, once the operation to train the untrained neural network has been initiated, the training operation can be performed within the memory array 130 without additional commands from the logic circuit system 122 and / or the host 102.

[0036] The control circuitry system 120 may additionally include a memory resource 124, which may be communicatively coupled to the logic circuitry system 122. The memory resource 124 may include volatile memory resources, non-volatile memory resources, or a combination of volatile and non-volatile memory resources. In some embodiments, the memory resource may be random access memory (RAM), such as static random access memory (SRAM). However, embodiments are not limited thereto, and the memory resource may be a cache memory, one or more registers, NVRAM, ReRAM, FeRAM, MRAM, PCM, "emerging" memory devices such as variable resistance memory resources, phase-change memory devices, memory devices including an array of selectable memory cells, or combinations thereof. In some embodiments, the memory resource 124 may serve as a cache memory for the logic circuitry system 122.

[0037] like Figure 1As shown, sensing circuitry 150 is coupled to memory array 130 and control circuitry 120. Sensing circuitry 150 may include one or more sensing amplifiers and one or more computing components. Sensing circuitry 150 provides additional storage space for memory array 130 and can sense (e.g., read, store, cache) data values ​​present in memory device 104. In some embodiments, sensing circuitry 150 may be located in a peripheral region of memory device 104. For example, sensing circuitry 150 may be located in a region of memory device 104 that is physically different from memory array 130. As described herein, sensing circuitry 150 may include sensing amplifiers, latches, flip-flops, etc., that can be configured to store data values. In some embodiments, sensing circuitry 150 may be provided as a register or a series of registers and may include the same number of storage locations (e.g., sensing amplifiers, latches, etc.) as the number of rows or columns of memory array 130 present. For example, if the memory array 130 contains approximately 16K rows or columns, then the sensing circuitry system 150 may contain approximately 16K storage locations. Therefore, in some embodiments, the sensing circuitry system 150 may be a register configured to hold up to 16K data values, but the embodiments are not limited thereto.

[0038] A peripheral sense amplifier (“PSA”) 170 may be coupled to the memory array 130, the sensing circuitry system 150, and / or the control circuitry system 120. The peripheral sense amplifier 170 provides additional storage space for the memory array 130 and can sense (e.g., read, store, cache) data values ​​present in the memory device 104. In some embodiments, the peripheral sense amplifier 170 may be located in a peripheral region of the memory device 104. For example, the peripheral sense amplifier 170 may be located in a region of the memory device 104 that is physically different from the memory array 130. As described herein, the peripheral sense amplifier 170 may include sense amplifiers, latches, triggers, etc., that can be configured to store data values. In some embodiments, the peripheral sense amplifier 170 may be provided in the form of a register or a series of registers and may include the same number of storage locations (e.g., sense amplifiers, latches, etc.) as the rows or columns of the memory array 130. For example, if the memory array 130 contains approximately 16K rows or columns, then the peripheral sense amplifier 170 may contain approximately 16K storage locations.

[0039] The peripheral sensing amplifier 170 can be used in conjunction with the sensing circuitry 150 and / or the memory array 130 to facilitate the execution of the neural network training operations described herein. For example, in some embodiments, the peripheral sensing amplifier 170 can store the neural network (e.g., in conjunction with the neural network described herein). Figure 2A and 2BThe neural networks 225 and 227 described herein contain portions and / or stored commands (e.g., PIM commands) to facilitate the execution of neural network training operations performed within the memory device 104.

[0040] Figure 1 Embodiments may include additional circuitry not described to avoid obscuring the embodiments of this disclosure. For example, memory device 104 may include an address circuitry that latches address signals provided on I / O connections via an I / O circuitry system. The address signals may be received and decoded by row decoders and column decoders to access memory device 104 and / or memory array 130. Those skilled in the art will understand that the number of address input connections may depend on the density and architecture of memory device 104 and / or memory array 130.

[0041] Figure 2A This is a functional block diagram of a memory device 204 comprising a control circuit system 220 and multiple memory banks 221-0 to 221-N for storing neural networks 225 / 227. The control circuit system 220, memory banks 221-0 to 221-N, and / or neural networks 225 / 227 may be referred to individually or together as a device. As used herein, "device" may refer to, but is not limited to, any of a variety of structures or combinations thereof, such as a circuit or circuit system, one or more dies, one or more modules, one or more devices, or one or more systems. Memory device 204 may be similar to... Figure 1 The memory device 104 described herein, and the control circuit system 220 may be similar to... Figure 1 The control circuit system 120 described herein.

[0042] The control circuitry system 220 can allocate multiple locations in the array of each respective memory bank 221-0 to 221-N to store library commands, application program instructions (e.g., for operation sequences), and variable parameters (e.g., in-memory processing (PIM) commands) associated with the operation of the memory device 204 for each memory bank 221-0 to 221-N. The control circuitry system 220 can send commands (e.g., PIM commands) whose program instructions are stored in a given memory bank 221-0 to 221-N to the multiple memory banks 221-0 to 221-N. As used herein, this contrasts with methods that cause operation via external processing components (e.g., ...). Figure 1 The host 102 described herein performs normal DRAM commands (e.g., read / write commands) on data. "PIM commands" are generated by processing elements (e.g., via...) within memory banks 221-0 to 221-N. Figure 1 The commands executed by the sensing circuit system 150 described herein. Therefore, the PIM command can correspond to a command that performs an operation within the memory banks 221-1 to 221-N without interfering with the host.

[0043] In some embodiments, a PIM command may be executed within the memory device 204 to store a trained neural network (e.g., neural network 225) in one of the memory banks (e.g., memory bank 221-0), to store an untrained neural network (e.g., neural network 227) in a different memory bank (e.g., memory bank 221-4), and / or to cause an operation to be performed to train an untrained neural network using the trained neural network.

[0044] As mentioned above, neural networks 225 and / or 227 can be trained over time using the input dataset to improve the accuracy of neural networks 225 / 227. However, in some embodiments, at least one of the neural networks (e.g., neural network 225) may be trained before being stored in one of the memory banks 221-0 to 221-N. In such embodiments, another neural network (e.g., neural network 227) may not have been trained before being stored in the memory banks 221-0 to 221-N. Continuing this example, an untrained neural network (e.g., neural network 227) can be trained using a trained neural network (e.g., neural network 225).

[0045] Memory libraries 221-0 to 221-N can be communicatively coupled via bus 229 (e.g., library-to-library transfer bus, communication subsystem, etc.). Bus 229 facilitates data and / or command transfers between memory libraries 221-0 to 221-N. In some embodiments, bus 229 facilitates data and / or command transfers between memory libraries 221-0 to 221-N as part of performing operations that train an untrained neural network (e.g., neural network 227) using a trained neural network (e.g., neural network 225).

[0046] Figure 2B This is another functional block diagram in the form of a control circuit system 220 and a memory device 221-0 to 221-N storing multiple neural networks 225-1 to 225-N and 227-1 to 227-M. The control circuit system 220, the multiple memory banks 221-0 to 221-N, and the neural networks 225 and 227 can be similar to... Figure 2A The control circuit system 220, multiple memory banks 221-0 to 221-N, and neural networks 225 and 227 are described herein.

[0047] In some embodiments, a corresponding trained neural network (e.g., neural networks 225-1 to 225-N) can perform the operation of training a corresponding untrained neural network (e.g., neural networks 227-1 to 227-M). For example, an untrained neural network 227-1 can be trained by a trained neural network 225-1, an untrained neural network 227-2 can be trained by a trained neural network 225-2, an untrained neural network 227-3 can be trained by a trained neural network 225-3, and / or an untrained neural network 227-M can be trained by a trained neural network 225-N, as described elsewhere herein.

[0048] Untrained neural networks (e.g., neural networks 227-1 to 227-M) can be trained substantially concurrently (e.g., in parallel) by trained neural networks (e.g., neural networks 225-1 to 225-N). As used herein, the term "substantially" means that the characteristic does not need to be absolute, but is close enough to achieve the advantages of the stated characteristic. For example, "substantially simultaneous" is not limited to performing operations at absolute simultaneousity and can include timing that is intended to be simultaneous but may not be exactly simultaneous due to manufacturing limitations. For example, due to read / write latency that may be presented by various interfaces and / or buses, training operations on untrained neural networks that are performed "substantially simultaneously" may not start or end at exactly the same time. However, at least one of the first untrained neural network (e.g., neural network 227-1) and the second untrained neural network (e.g., neural network 227-2) can be trained by corresponding trained neural networks (e.g., neural networks 225-1 and 225-2) such that training operations are performed simultaneously without regard to whether training operations for the first and second untrained neural networks start or terminate before other training operations. However, the embodiments are not limited thereto, and in some embodiments, an untrained neural network (e.g., neural networks 227-1 to 227-M) can be trained by training neural networks 225-1 to 225-N.

[0049] In some embodiments, trained neural networks (e.g., neural networks 225-1 to 225-N) and / or untrained neural networks (e.g., neural networks 227-1 to 227-M) may be portions or subsets of a larger neural network. For example, a trained or untrained neural network may be divided into smaller components and stored on multiple libraries 221 of the memory device 204. In some embodiments, the control circuitry system 220 may control the operation of splitting the entire neural network into components or subsets. By allowing the neural network to be split into smaller components or subsets, the storage and / or training of the neural network can be achieved within the storage limits of the memory device 204, which includes multiple memory libraries 221-0 to 221-N.

[0050] In a non-limiting example, the system may include a memory device 204 comprising eight memory banks 221-0 to 220-N. The system may additionally include a control circuitry system 220 residing on the memory device 204 and communicatively coupled to the eight memory banks 221-0 to 221-N. As used herein, the term "residing on" means that something is physically located on a particular component. For example, "residing on" the control circuitry system 220 means that the hardware circuitry system including the control circuitry system 220 is physically located on the memory device 204. The term "residing on" may be used interchangeably herein with other terms such as "deployed on" or "located on".

[0051] The control circuitry system 220 can control the storage of four distinct trained neural networks (e.g., neural networks 225-1, 225-2, 225-3, and 225-N) in four memory banks (e.g., memory banks 221-0, 221-1, 221-2, and 221-3). The control circuitry system 220 can also control the storage of four distinct untrained neural networks (e.g., neural networks 227-1, 227-2, 227-3, and 227-N) in four different memory banks (e.g., memory banks 221-4, 221-5, 221-6, and 221-N), such that each of the eight memory banks 221-0 to 221-N stores either a trained neural network or an untrained neural network. In some embodiments, at least two of the trained neural networks and / or at least two of the untrained neural networks may be neural networks of different types.

[0052] For example, at least two of the trained neural networks and / or at least two of the untrained neural networks may be feedforward neural networks or backpropagation neural networks. However, the embodiments are not limited to this, and at least two of the trained neural networks and / or at least two of the untrained neural networks may be perceptron neural networks, radial basis function neural networks, deep feedforward neural networks, recurrent neural networks, long-term / short-term memory neural networks, gated recurrent unit neural networks, autoencoder (AE) neural networks, variational AE neural networks, denoising AE neural networks, sparse AE neural networks, Markov chain neural networks, Hopfield neural networks, Boltzmann machine (BM) neural networks, restricted BM neural networks, deep belief neural networks, deep convolutional neural networks, deconvolutional neural networks, deep convolutional inverse graph neural networks, generative adversarial neural networks, liquid machine neural networks, extreme learning machine neural networks, echo state neural networks, deep residual neural networks, Kohonen neural networks, support vector machine neural networks, and / or neural Turing machine neural networks, etc.

[0053] In some embodiments, the control circuitry 220 can control the execution of multiple neural network training operations to train an untrained neural network using a trained neural network, without signaling generated by circuitry outside the memory device 204. Compared to methods that do not perform neural network training within the memory device 204, performing neural network training without signaling generated by circuitry outside the memory device 204 (e.g., performing neural network training within the memory device 204 or "on-chip") reduces data movement in and out of the memory device 204. This allows for reduced power consumption during neural network training operations and / or reduced load on the host computing system (e.g., ...). Figure 1 The dependency of host 102 as described in the document. In addition, neural network training can be automated, which can reduce the amount of time required to train the neural network.

[0054] As used herein, a “neural network training operation” includes operations performed to determine at least one hidden layer in a neural network. Generally, a neural network may include at least one input layer, at least one hidden layer, and at least one output layer. Each layer may contain multiple neurons capable of receiving inputs and producing weighted outputs. In some embodiments, neurons in the hidden layer may calculate a weighted sum and / or average of the inputs received from the input layer and their respective weights, and transmit this information to the output layer.

[0055] In some embodiments, neural network training operations can be performed by utilizing the knowledge learned by the trained neural network during its training of the untrained neural network. This reduces the amount of time and resources spent training the untrained neural network by minimizing the retraining of information already learned by the trained neural network. Additionally, the embodiments described herein can allow a neural network trained under a particular training method to train an untrained neural network using different training methods. For example, a neural network can be trained using the TensorFlow method and subsequently trained as an untrained neural network using the MobileNet method (and vice versa). However, the embodiments are not limited to these specific examples, and other training methods are contemplated within the scope of this disclosure.

[0056] As described above, in some embodiments, the control circuit system 220 can control the execution of multiple neural network training operations so that the multiple neural network training operations can be executed in substantially parallel.

[0057] In some embodiments, the control circuitry 220 may cause an operation to convert data associated with a neural network (e.g., a trained neural network and / or an untrained neural network) from one data type to another before the trained and / or untrained neural network is stored in memory banks 221-0 to 221-N and / or before the neural network is transferred to a circuitry outside the memory device 204. As used herein, “data type” generally refers to the format in which data is stored. Non-limiting examples of data types include IEEE 754 floating-point format, fixed-point binary format, and / or universal number (unum) format, such as Type III unum and / or posit. Therefore, in some embodiments, the control circuitry 220 may cause an operation to convert data associated with a neural network (e.g., a trained neural network and / or an untrained neural network) from a floating-point or fixed-point binary format to a universal number or posit format before the trained and / or untrained neural network is stored in memory banks 221-0 to 221-N and / or before the neural network is transferred to a circuitry outside the memory device 204.

[0058] Compared to IEEE 754 floating-point or fixed-point binary formats, which contain subsets of sign, mantissa, and exponent bits, hypothetical general-purpose number formats, such as those containing subsets of sign, state, mantissa, and exponent bits, allow hypothetical numbers to have greater accuracy, precision, and / or dynamic range than floating-point numbers or other numeric formats. Furthermore, hypothetical numbers can reduce or eliminate overflow, underflow, NaN, and / or other extreme cases associated with floating-point numbers and other numeric formats. Additionally, using hypothetical numbers allows for the use of fewer bits to represent numerical values ​​(e.g., numbers) compared to floating-point numbers or other numeric formats.

[0059] In some embodiments, control circuitry 220 may determine that at least one of the untrained neural networks has been trained and cause the trained neural network to be transferred to a circuitry outside of memory device 204. Additionally, in some embodiments, control circuitry 220 may determine that at least one of the untrained neural networks has been trained and cause an operation to modify the accuracy, dynamic range, or both of the information (e.g., data) associated with the trained neural network. However, embodiments are not limited thereto, and in some embodiments, control circuitry 220 may cause an operation to modify the accuracy, dynamic range, or both of the information (e.g., data) associated with the trained or untrained neural network before the trained or untrained neural network is stored in memory banks 221-0 to 221-N.

[0060] As used herein, “precision” refers to the number of bits in a bit string used to perform a computation. For example, if every bit in a 16-bit bit string is used when performing a computation, the bit string can be said to have 16-bit precision. However, if only 8 bits of a 16-bit bit string are used when performing a computation (e.g., if the first 8 bits of the bit string are zero), the bit string can be said to have 8-bit precision. As the precision of the bit string increases, computations can be performed with higher accuracy. Conversely, as the precision of the bit string decreases, computations can be performed with lower accuracy. For example, an 8-bit bit string may correspond to a data range consisting of 255 (256) precision steps, while a 16-bit bit string may correspond to a data range consisting of 65,536 (63,536) precision steps.

[0061] As used herein, "dynamic range" or "dynamic range of data" refers to the ratio between the maximum and minimum values ​​that can be represented by a bit string with a specific precision associated with it. For example, the maximum numerical value that can be represented by a bit string with a specific precision associated with it can determine the dynamic range of the bit string's data format. For a bit string in a general number (e.g., hypothetical number) format, the dynamic range can be determined by the numerical values ​​of a subset of the exponent bits of the bit string.

[0062] Dynamic range and / or accuracy can have associated variable range thresholds. For example, the dynamic range of data can correspond to the applications that use the data and / or the various computations that use the data. This may be due to the fact that one application may expect a different dynamic range than another application, and / or because some computations may require different data dynamic ranges. Therefore, the embodiments herein can allow the dynamic range of data to be changed to suit the requirements of different applications and / or computations. In contrast to methods that do not allow manipulation of the dynamic range of data to suit the requirements of different applications and / or computations, the embodiments herein can improve resource utilization and / or data accuracy by allowing the dynamic range of data to be changed based on the applications and / or computations that use the data.

[0063] Figure 3 This is another functional block diagram of a memory device 304 comprising a control circuit system 320 and multiple memory banks 321-0 to 321-N storing neural networks 325 / 327. The control circuit system 320, memory banks 321-0 to 321-N, and / or neural networks 325 / 227 may be referred to individually or together as a device. As used herein, "device" may refer to, but is not limited to, any of a variety of structures or combinations thereof, such as a circuit or circuit system, one or more dies, one or more modules, one or more devices, or one or more systems. Memory device 304 may be similar to... Figure 2A and 2B The memory device 204 described herein, and the control circuit system 320 may be similar to... Figure 2A and 2B The control circuit system 220 is described herein. In this document, memory banks 321-0 to 321-N can be similar to... Figure 2A and 2B The memory banks 221-0 to 221-N described herein, and the neural network 325 can be similar to... Figure 2A and 2B The neural network 225 described herein, and the neural network 327 may be similar to Figure 2A and 2B The neural network 227 is described in the text. Although... Figure 3 Not explicitly shown, but memory banks 321-0 to 321-N can be accessed via a bus (e.g., as shown in this document). Figure 2A and 2B The buses 229 described herein are coupled to each other in a communicative manner.

[0064] like Figure 3As shown, neural networks 325 and 327 can be distributed across multiple memory libraries 321 of the memory device 304. For example, a first subset of the memory libraries (e.g., memory libraries 321-0 to 321-3) can be configured to store the neural network 325, and a second subset of the memory libraries (e.g., memory libraries 321-4 to 321-N) can be configured to store the neural network 327. For example, in some embodiments, the first subset may include half of the total number of memory libraries 321-0 to 321-N associated with the memory device 304, and the second subset may include the other half of the total number of memory libraries 321-0 to 321-N associated with the memory device 304. However, embodiments are not limited to this particular configuration, and the memory libraries 321 may be divided into more than two subsets and / or subsets may contain more than four memory libraries 321 and / or fewer than four memory libraries 321.

[0065] In a non-limiting example, the device may include a memory device 304 comprising a plurality of memory cell libraries 321-0 to 321-N and a control circuitry 320 residing on the memory device 304 and communicatively coupled to each of the plurality of memory libraries 321-0 to 321-N. In some embodiments, the control circuitry 320 may control the storage of a first neural network (e.g., neural network 325) in a first subset of the plurality of memory libraries 321-0 to 321-N (e.g., memory libraries 321-0 to 321-3). The control circuitry 320 may further control the storage of a second neural network (e.g., neural network 327) in a second subset of the plurality of memory libraries 321-0 to 321-N (e.g., memory libraries 321-4 to 321-N) and / or control the execution of neural network training operations such that the second neural network is trained by the first neural network. However, the embodiments are not limited thereto, and in some embodiments, the control circuit system 320 can control the storage of the third neural network in a third subset of a plurality of memory libraries and can control the execution of neural network training operations such that the third neural network is trained by the first neural network and / or the second neural network.

[0066] The first neural network may be trained before being stored in a first subset of multiple memory libraries 321-0 to 321-N, and the second neural network may not be trained before being stored in a second subset of multiple memory libraries 321-0 to 321-N (e.g., the second neural network may be untrained). Therefore, in some embodiments, the second neural network can be trained by the first neural network.

[0067] As described in more detail above, the control circuitry system 320 can control the storage of the first neural network, the storage of the second neural network, or the execution of neural network training operations, or any combination thereof, without signaling generated by components outside the memory device 304. For example, the storage of the first neural network, the storage of the second neural network, or the execution of neural network training operations, or any combination thereof, can be performed entirely within the memory device 304 without requiring a host external to the memory device 304 (e.g., [other components]). Figure 1 Additional inputs to the host 102 or other circuit systems as described herein.

[0068] Figure 4 This is a flowchart illustrating an example method 430 of a memory device for training a neural network, according to several embodiments of the present disclosure. Method 430 can be executed by processing logic, which may include hardware (e.g., processing means, circuitry, dedicated logic, programmable logic, microcode, device hardware, integrated circuits, etc.), software (e.g., instructions that run or execute on the processing means), or a combination thereof. Although shown in a specific order or sequence, the order of processes may be modified unless otherwise specified. Therefore, it should be understood that the illustrated embodiments are merely examples, and the illustrated processes may be performed in different orders, and some processes may be performed in parallel. Furthermore, one or more processes may be omitted in various embodiments. Therefore, not all processes are required in every embodiment. Other process flows are also possible.

[0069] At box 432, method 430 may include writing data associated with the input or output layer of the first neural network into a first memory bank of the memory device. The first memory bank may be similar to that described herein. Figure 2A and 2B The memory bank 221-0 to 221-N of the memory device 204 described herein, and the first neural network may be similar to that described herein. Figure 2A One of the neural networks 225 / 227 described in 2b.

[0070] At box 434, method 430 may include writing data associated with the input or output layer of the second neural network into a second memory bank of the memory device. The second memory bank may be similar to that described herein. Figure 2A and 2B The memory bank 221-0 to 221-N of the memory device 204 described herein, and the second neural network may be similar to that described herein. Figure 2A One of the neural networks 225 / 227 described in 2b. In some embodiments, the first neural network may be trained before being stored in the first memory bank, and the second neural network may not be trained before being stored in the second memory bank.

[0071] At block 436, method 430 may include determining one or more weights within a memory device for a hidden layer of a first neural network or a second neural network, or both. For example, method 430 may include performing a neural network training operation to train the first neural network or the second neural network by determining the weights of the hidden layers of at least one of the neural networks. Method 430 may further include performing a neural network training operation to train the other of the first neural network or the second neural network using a training set learned by one of the first neural networks or the second neural network.

[0072] As described above, in some embodiments, method 430 may include performing neural network training operations locally within a memory device. For example, method 430 may be included in a host computing system that is coupled to the memory device (e.g., herein...) without impeding... Figure 1 The neural network training operation is performed in the case of host 102 as described herein. In some embodiments, method 430 may include at least in part based on circuitry residing on a memory device (e.g., as described herein). Figure 2A and 2B The control circuit system 220 described herein generates control signals to perform neural network training operations.

[0073] In some embodiments, the first neural network may be a first type of neural network, and the second neural network may be a second type of neural network. For example, the first neural network may be a feedforward neural network and the second neural network may be a backpropagation neural network, or vice versa. However, the embodiments are not limited to this, and the first neural network and / or the second neural network may be a perceptron neural network, a radial basis function neural network, a deep feedforward neural network, a recurrent neural network, a long-term / short-term memory neural network, a gated recurrent unit neural network, an autoencoder (AE) neural network, a variational AE neural network, a denoising AE neural network, a sparse AE neural network, a Markov chain neural network, a Hopfield neural network, a Boltzmann machine (BM) neural network, a restricted BM neural network, a deep belief neural network, a deep convolutional neural network, a deconvolutional neural network, a deep convolutional inverse graph neural network, a generative adversarial neural network, a liquid machine neural network, an extreme learning machine neural network, an echo state neural network, a deep residual neural network, a Coholon neural network, a support vector machine neural network, and / or a neural Turing machine neural network, etc.

[0074] Figure 5This is a flowchart illustrating another example of a method 540 corresponding to a memory device for training a neural network according to several embodiments of the present disclosure. Method 540 can be executed by processing logic, which may include hardware (e.g., processing means, circuitry, dedicated logic, programmable logic, microcode, device hardware, integrated circuits, etc.), software (e.g., instructions that run or execute on the processing means), or a combination thereof. Although shown in a specific order or sequence, the order of processes may be modified unless otherwise specified. Therefore, it should be understood that the illustrated embodiments are merely examples, and the illustrated processes may be performed in different orders, and some processes may be performed in parallel. Furthermore, one or more processes may be omitted in various embodiments. Therefore, not all processes are required in every embodiment. Other process flows are also possible.

[0075] At block 542, method 540 may include storing multiple different neural networks in respective memory banks of a memory device. In some embodiments, at least one neural network may be trained and at least one neural network may be untrained. The multiple memory banks may be similar to those described herein. Figure 2A and 2B The memory bank 221-0 to 221-N of the memory device 204 described herein, and the neural network can be similar to that described herein. Figure 2A The neural networks 225 / 227 described in 2b.

[0076] At block 544, method 540 may include performing a neural network training operation to train at least one untrained neural network using at least one trained neural network. As described above, in some embodiments, method 430 may include performing the neural network training operation locally within a memory device.

[0077] In some embodiments, the memory device may include eight memory banks. For example, four trained neural networks may be stored in four corresponding memory banks of the memory device, and four untrained neural networks may be stored in four different corresponding memory banks of the memory device. In such embodiments, method 540 may include performing a neural network training operation using the corresponding trained neural networks to train the corresponding untrained neural networks within the memory device. In some embodiments, method 430 may include performing a neural network training operation to train the corresponding untrained neural networks in substantially parallel.

[0078] Method 540 may additionally include a control circuit system residing on the memory device (e.g., as described herein). Figure 2A and 2BThe control circuit system 220 described herein determines the completion of the neural network training operation and, in response to the signal generated by the control circuit system, transmits the neural network that has undergone the complete neural network training operation to the external memory device.

[0079] In some embodiments, method 540 may include performing operations by controlling the circuitry system to modify the accuracy, dynamic range, or both of the information associated with the neural network undergoing the complete training operation before transmitting the neural network, which has undergone a complete training operation, to a circuitry system outside the memory device.

[0080] Figure 6 This is a schematic diagram illustrating a portion of a memory array including a sensing circuitry system according to several embodiments of the present disclosure. Sensing component 650 indicates that it can correspond to... Figure 1 One of several sensing components in the sensing circuit system 150 shown in the figure.

[0081] exist Figure 6 In the example shown, memory array 630 is a DRAM array of transistor-capacitor (1T1C) memory cells, where the transistors act as access devices and the capacitors act as storage elements; however, other configuration embodiments can be used (e.g., a 2T2C configuration with two transistors and two capacitors per memory cell). In this example, the first memory cell includes transistor 651-1 and capacitor 652-1, and the second memory cell includes transistor 651-2 and capacitor 652-2, etc. In several embodiments, the memory cell can be a destructive read memory cell (e.g., reading data stored in the cell destroys the data, causing the data initially stored in the cell to be refreshed after being read).

[0082] The cells of the memory array 630 can be arranged in rows coupled via word lines 662-X (row X), 662-Y (row Y), etc., and via complementary sensing line pairs (e.g., Figure 6 The columns are coupled, with the digital line 653-1 labeled DIGIT(n) and the column labeled DIGIT(n)_. Although in Figure 6 Only one pair of complementary digital lines is shown in the illustration, but embodiments of this disclosure are not limited thereto, and the memory cell array may include additional columns of memory cells and digital lines (e.g., 4,096, 8,192, 16,384, etc.).

[0083] Memory cells can be coupled to different digital lines and word lines. For example, in this instance, the first source / drain region of transistor 651-1 is coupled to digital line 653-1, the second source / drain region of transistor 651-1 is coupled to capacitor 652-1, and the gate of transistor 651-1 is coupled to word line 662-Y. The first source / drain region of transistor 651-2 is coupled to digital line 653-2, the second source / drain region of transistor 651-2 is coupled to capacitor 652-2, and the gate of transistor 651-2 is coupled to word line 662-X. For example... Figure 6 The cell board shown can be coupled to each of capacitors 652-1 and 652-2. The cell board can be a common node that can be applied a reference voltage (e.g., ground) in various memory array configurations.

[0084] According to several embodiments of this disclosure, digital lines 653-1 and 653-2 of the memory array 630 are coupled to a sensing component 650. In this example, the sensing component 650 includes a sensing amplifier 654 corresponding to a respective column of memory cells (e.g., coupled to a corresponding pair of complementary digital lines) and a computing component 665. The sensing amplifier 654 is coupled to the pair of complementary digital lines 653-1 and 653-2. The computing component 665 is coupled to the sensing amplifier 654 via gates 655-1 and 655-2. The gates of gates 655-1 and 655-2 are coupled to logic operation selection logic 613.

[0085] Selection logic 613 may include pass-gate logic for controlling pass-through gates and switch-gate logic for controlling switch gates, the pass-through gates being coupled to untransposed complementary digital line pairs between sense amplifier 654 and computing component 665, and the switch gates being coupled to transposed complementary digital line pairs between sense amplifier 654 and computing component 665. Selection logic 613 may be coupled to the pairs of complementary digital lines 653-1 and 653-2 and configured to perform logical operations on data stored in array 630. For example, selection logic 613 may be configured to control the continuity (e.g., on / off) of pass-through gates 655-1 and 655-2 based on the selected logical operation performed.

[0086] An operable sense amplifier 654 determines a data value (e.g., a logic state) stored in a selected memory cell. The sense amplifier 654 may include a cross-coupled latch 615 (e.g., the gates of one pair of transistors (e.g., n-channel transistors 661-1 and 661-2) are cross-coupled to the gates of another pair of transistors (e.g., p-channel transistors 629-1 and 629-2), which may be referred to herein as a master latch. However, embodiments are not limited to this example.

[0087] In operation, when a memory cell is being sensed (e.g., read), the voltage on one of digital lines 653-1 or 653-2 will be slightly greater than the voltage on the other of digital lines 653-1 or 653-2. The ACT signal and the RNL* signal can be driven low to enable (e.g., activate) the sense amplifier 654. The digital line 653-1 or 653-2 with the lower voltage will turn on one of transistors 629-1 or 629-2 to a greater extent than the other of transistors 629-1 or 629-2, thereby driving the digital line 654-1 or 654-2 with the higher voltage high to a greater extent than driving the other digital line 654-1 or 654-2 high.

[0088] Similarly, the digital line 654-1 or 654-2 with the higher voltage will turn on one of the transistors 661-1 or 661-2 to a greater extent than the other, thereby driving the digital line 654-1 or 654-2 with the lower voltage to a greater extent than driving the other digital line 654-1 or 654-2 to a higher level. Therefore, after a short delay, the digital line 654-1 or 654-2 with the slightly higher voltage is driven through the supply voltage V of the source transistor. CC The voltage across the reference voltage (e.g., ground) of the absorption transistor is also driven, and another digital line 654-1 or 654-2 is driven to pass through the absorption transistor. Therefore, the cross-coupled transistors 661-1 and 661-2 and transistors 629-1 and 629-2 act as a sense amplifier pair, amplifying the differential voltage on digital lines 654-1 and 654-2 and using it to latch the data value sensed from a selected memory cell.

[0089] Examples are not limited to Figure 6 The sensing component configuration described herein. As an example, the sensing amplifier 654 may be a current-mode sensing amplifier and / or a single-ended sensing amplifier (e.g., a sensing amplifier coupled to a digital line). Furthermore, embodiments of this disclosure are not limited to, for example... Figure 6 The folded digital line architecture shown.

[0090] Sensing component 650 may be one of a plurality of sensing components selectively coupled to a shared I / O line. Therefore, sensing component 650 may be used in association with data inverted and stored in memory according to several embodiments of the present disclosure.

[0091] In this example, the sense amplifier 654 includes a balanced circuit system 659, which can be configured to balance digital lines 654-1 and 654-2. The balanced circuit system 659 includes transistors 658 coupled between digital lines 654-1 and 654-2. The balanced circuit system 659 also includes transistors 658, each having a voltage coupled to a balanced voltage (e.g., V).DD / 2) Transistors 656-1 and 656-2 in the first source / drain region, where V DD This is the supply voltage associated with the array. The second source / drain region of transistor 656-1 is coupled to digital line 654-1, and the second source / drain region of transistor 656-2 is coupled to digital line 654-2. The gates of transistors 658, 656-1, and 656-2 may be coupled together and coupled to the balancing (EQ) control signal line 657. Therefore, activating EQ enables transistors 658, 656-1, and 656-2, which effectively shorts digital lines 654-1 and 654-2 together and to the balancing voltage (e.g., V). DD / 2). Although Figure 6 The illustration shows that the sensing amplifier 654 includes a balancing circuit system 659, but embodiments are not so limited, and the balancing circuit system 659 may be implemented separately from the sensing amplifier 654, or implemented in conjunction with... Figure 6 The configurations shown may be implemented in different configurations or not at all.

[0092] like Figure 6 As shown, computing component 665 may also include a latch, referred to herein as secondary latch 664. Secondary latch 664 may be configured and operate in a manner similar to that described above with respect to primary latch 663, except that the cross-coupled p-channel transistors (e.g., PMOS transistors) included in the secondary latch can couple their respective sources to a supply voltage 612-2 (e.g., V). DD Furthermore, the cross-coupled n-channel transistors (e.g., NMOS transistors) of the secondary latch can selectively couple their respective sources to a reference voltage 612-1 (e.g., ground), thereby continuously enabling the secondary latch. The configuration of the computing component 665 is not limited to... Figure 6 The configuration shown in the document is feasible, and various other embodiments are also possible.

[0093] In some embodiments, as described above, the sensing component 650 may be combined as described above to train data stored in a memory bank (e.g., as described herein). Figure 2A and 2B The neural network in the memory library 221 described herein (e.g., the one described in this paper) Figure 2A and 2B The operation is performed by executing one or more operations of the neural network 225 and / or 227 described above. For example, as part of performing the training operation described above, data associated with the neural network and / or the training of the neural network may be processed or manipulated on the sensing component 650.

[0094] While specific embodiments have been shown and described herein, those skilled in the art will understand that arrangements calculated to achieve the same results may replace the specific embodiments shown. This disclosure is intended to cover modifications or variations of one or more embodiments of this disclosure. It should be understood that the above description has been carried out illustratively and not restrictively. Combinations of the above embodiments and other embodiments not specifically described herein will be apparent to those skilled in the art upon review of the above description. The scope of one or more embodiments of this disclosure includes other applications in which the above structures and processes are used. Therefore, the scope of one or more embodiments of this disclosure should be determined with reference to the appended claims together with the full scope of the equivalents given by such claims.

[0095] In the foregoing detailed embodiments, some features are grouped together in a single embodiment for the purpose of simplifying this disclosure. This approach of the disclosure should not be construed as reflecting an intention that the disclosed embodiments must use more features than expressly stated in each claim. In fact, as reflected in the appended claims, the subject matter of the invention lies in less than all the features of a single disclosed embodiment. Therefore, the appended claims are hereby incorporated into the detailed embodiments, wherein each claim is, in itself, a separate embodiment.

Claims

1. A method for training a neural network for a memory device, comprising: The first neural network is broken down into its components; Data associated with the input or output layer of the first neural network is written across a first plurality of volatile memory banks of a dynamic random access memory (DRAM) device and within a plurality of peripheral sense amplifiers coupled to the volatile memory banks, such that each of the first plurality of volatile memory banks stores at least one component of the first neural network and the plurality of peripheral sense amplifiers store at least one component of the first neural network. The second neural network is broken down into its components; Data associated with the input or output layer of the second neural network is written across a second plurality of volatile memory libraries of the DRAM device and within the plurality of peripheral sense amplifiers, such that each of the second plurality of volatile memory libraries stores at least one component of the second neural network and the plurality of peripheral sense amplifiers store at least one component of the second neural network. and Determine one or more weights within the DRAM device for the hidden layer of the first neural network or the second neural network or both.

2. The method of claim 1, wherein the components of the first neural network are trained before being stored in the first plurality of memory banks and the plurality of peripheral sensing amplifiers, and wherein the components of the second neural network are not trained before being stored in the second plurality of memory banks and the plurality of peripheral sensing amplifiers.

3. The method of claim 1, wherein, as part of a neural network training operation, the determination of one or more weights of the hidden layer of the first neural network or the second neural network or both is performed, and wherein the method further comprises performing the neural network training operation to train the other of the components of the first neural network or the second neural network using a training set learned by one of the components of the first neural network or the second neural network, such that each neural network is trained by a neural network trained on a different training set.

4. The method according to any one of claims 1 to 3, wherein, as part of a neural network training operation, the determination of one or more weights of the hidden layer of the first neural network or the second neural network or both is performed, and wherein the method further includes performing the neural network training operation locally within the memory device.

5. The method according to any one of claims 1 to 3, wherein, as part of a neural network training operation, the determination of one or more weights of the hidden layer of the first neural network or the second neural network or both is performed, and wherein the method further includes performing the neural network training operation without interfering with a host computing system that can be coupled to the memory device.

6. The method according to any one of claims 1 to 3, wherein, as part of a neural network training operation, the determination of one or more weights of the hidden layer of the first neural network or the second neural network or both is performed, and wherein the method further comprises performing the neural network training operation based at least in part on control signaling generated by a circuit system residing on the memory device.

7. The method according to any one of claims 1 to 3, wherein the first neural network is a first type of neural network, and wherein the second neural network is a second type of neural network.

8. An apparatus for training a neural network using a memory device, comprising: A volatile memory device comprising a library of multiple dynamic random access memory (DRAM) cells; and A control circuitry system, residing on the volatile memory device and communicatively coupled to each of the plurality of DRAM cell libraries, wherein the control circuitry system performs the following operations: The control system breaks down the first neural network into its components; The operation of controlling the writing of data associated with the input or output layer of the first neural network across a plurality of DRAM cell libraries and within a plurality of peripheral sense amplifiers coupled to the plurality of DRAM cell libraries, such that each of the plurality of libraries stores at least one component of the first neural network and the plurality of peripheral sense amplifiers stores at least one component of the first neural network. The control breaks down the second neural network into its components; The operation of controlling the writing of data associated with the input or output layer of the second neural network across a second plurality of DRAM cell libraries and within the plurality of peripheral sense amplifiers, such that each of the second plurality of libraries stores at least one component of the second neural network and the plurality of peripheral sense amplifiers stores at least one component of the second neural network; and The execution of the neural network training operation is controlled to cause the second neural network to be trained by the first neural network by determining one or more weights of the hidden layer of the second neural network.

9. The device of claim 8, wherein the first neural network is trained prior to being stored in the first plurality of libraries in the plurality of DRAM cell libraries and in the plurality of peripheral sensing amplifiers.

10. The device according to any one of claims 8 to 9, wherein the second neural network was not trained prior to being stored in the second plurality of libraries in the plurality of DRAM cell libraries and within the plurality of peripheral sensing amplifiers.

11. The device according to any one of claims 8 to 9, wherein the control circuitry system controls the writing of data associated with the input layer or the output layer of the first neural network, the writing of data associated with the input layer or the output layer of the second neural network, or the execution of the neural network training operation, or any combination thereof, without signaling generated by components outside the memory device.

12. The device according to any one of claims 8 to 9, wherein the control circuit system performs the following operations: Controlling the operation of writing data associated with the input or output layer of the third neural network into a third plurality of libraries within the plurality of DRAM cell libraries; and The execution of the neural network training operation is controlled to cause the third neural network to be trained by the first neural network, the second neural network, or both by determining one or more weights of the hidden layer of the third neural network.

13. The device according to any one of claims 8 to 9, wherein the first plurality of libraries comprises half of the total number of memory libraries associated with the memory device, and the second plurality of libraries comprises the other half of the total number of memory libraries associated with the memory device.

14. A system for training a neural network using a memory device, comprising: A volatile memory device comprising eight dynamic random access memory (DRAM) libraries; and A control circuitry system, communicatively coupled to the eight DRAM libraries, wherein the control circuitry system performs the following operations: Control breaks down trained or untrained neural networks into components; The operation of controlling the writing of data associated with the input or output layers of four dissimilar trained neural networks into four of the eight DRAM libraries and into a plurality of peripheral sense amplifiers coupled to the eight DRAM libraries is such that each of the four DRAM libraries stores at least one component of the four dissimilar trained neural networks and the plurality of peripheral sense amplifiers stores at least one component of the four dissimilar trained neural networks. The operation of controlling the writing of data associated with the input or output layers of four dissimilar untrained neural networks into four different DRAM libraries out of the eight DRAM libraries and into the plurality of peripheral sense amplifiers, such that each of the four different DRAM libraries stores at least one component of the four dissimilar untrained neural networks and the plurality of peripheral sense amplifiers stores at least one component of the four dissimilar untrained neural networks. and Without signaling generated by circuitry outside the volatile memory device, the execution of multiple neural network training operations is controlled to cause the four dissimilar untrained neural networks to be trained by the four dissimilar trained neural networks by determining one or more weights of the hidden layers of the four dissimilar untrained neural networks.

15. The system of claim 14, wherein the control circuitry controls the execution of the plurality of neural network training operations such that the plurality of neural network training operations are executed substantially in parallel.

16. The system of claim 14, wherein the control circuitry performs the following operations: Determine that at least one of the four dissimilar untrained neural networks has been trained; and This causes the execution of operations to alter the accuracy or dynamic range, or both, of the information associated with the already trained neural network.

17. The system of claim 14, wherein the control circuitry performs the following operations: Determine that at least one of the four dissimilar untrained neural networks has been trained; and This causes the trained neural network to be transmitted to a circuit system outside the memory device.

18. The system according to any one of claims 14 to 17, wherein at least two of the four dissimilar trained neural networks or at least two of the four dissimilar untrained neural networks or both are neural networks of different types.

19. The system according to any one of claims 14 to 17, wherein the control circuitry controls the execution of the plurality of neural network training operations by determining one or more weights of the hidden layer of at least one of the four dissimilar untrained neural networks.

20. The system according to any one of claims 14 to 17, wherein the control circuitry resides on the volatile memory device.