WO2021072817A1 - 一种基于忆阻交叉阵列的机器人控制系统及方法 - Google Patents

一种基于忆阻交叉阵列的机器人控制系统及方法 Download PDF

Info

Publication number
WO2021072817A1
WO2021072817A1 PCT/CN2019/114958 CN2019114958W WO2021072817A1 WO 2021072817 A1 WO2021072817 A1 WO 2021072817A1 CN 2019114958 W CN2019114958 W CN 2019114958W WO 2021072817 A1 WO2021072817 A1 WO 2021072817A1
Authority
WO
WIPO (PCT)
Prior art keywords
memristor
circuit
robot
cross array
array
Prior art date
Application number
PCT/CN2019/114958
Other languages
English (en)
French (fr)
Inventor
缪峰
梁世军
王聪
杨再正
Original Assignee
南京大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 南京大学 filed Critical 南京大学
Priority to US17/763,495 priority Critical patent/US20220331952A1/en
Publication of WO2021072817A1 publication Critical patent/WO2021072817A1/zh

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00—Programme-controlled manipulators
    • B25J9/16—Programme controls
    • B25J9/1602—Programme controls characterised by the control system, structure, architecture
    • B25J9/161—Hardware, e.g. neural networks, fuzzy logic, interfaces, processor
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J13/00—Controls for manipulators
    • B25J13/08—Controls for manipulators by means of sensing devices, e.g. viewing or touching devices
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00—Programme-controlled manipulators
    • B25J9/16—Programme controls
    • B25J9/1628—Programme controls characterised by the control loop
    • B25J9/163—Programme controls characterised by the control loop learning, adaptive, model based, rule based expert control
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00—Programme-controlled manipulators
    • B25J9/16—Programme controls
    • B25J9/1656—Programme controls characterised by programming, planning systems for manipulators
    • B25J9/1664—Programme controls characterised by programming, planning systems for manipulators characterised by motion, path, trajectory planning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G06N3/065—Analogue means
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning

Definitions

  • the invention relates to the field of automatic control of memristive cross-arrays, in particular to a robot control system and method based on memristive cross-arrays.
  • the existing intelligent robot technology is based on a digital control circuit with a digital processor as the core in hardware, and is based on various neural network algorithms to realize intelligent applications in software. As this kind of robot driven by neural network algorithm, it requires low power consumption and fast processing speed. On current application platforms such as robots, relying on the existing digital control circuit to drive the neural network algorithm brings severe challenges to its hardware. Challenge.
  • the information control flow is as described below.
  • the physical quantities in the environment are collected and converted into digital signals, which are transmitted to the digital processor through digital communication.
  • the digital processor the information is processed according to the neural network algorithm.
  • the neural network algorithm In the general neural network algorithm model, it is composed of several layers of similar processing procedures. Each layer of the network contains a large number of weight parameters.
  • the digital information and network weight parameters in each layer of the network are subjected to matrix multiplication and nonlinear operations to obtain the output of this layer, which is further transformed into the input of the next layer.
  • the input information is transformed into the calculation result, and then the control signal is modulated according to the output result to control various intelligent actions of the robot.
  • the neural network needs to be trained with certain feedback.
  • the most common methods are error back propagation and gradient descent methods, and these training methods involve a large number of parameter adjustments and calculations.
  • the core responsible for information processing mainly designs the digital controller circuit.
  • the integrated digital chips it contains are all logic circuits based on complementary metal oxide semiconductor technology, using transistors as basic components, and they mainly include circuit modules such as general-purpose processors and memory.
  • circuit modules such as general-purpose processors and memory.
  • the sensing part the sensor is often made through physical effects, and the components are affected by physical quantities in the environment and have different electrical performances. This response is reflected in electrical analog signals.
  • the sensor element converts the analog quantity in the environment into the electrical analog quantity, it also needs to cooperate with the analog-to-digital converter to convert it into a digital signal, which is transmitted to the subsequent control circuit in the form of communication for subsequent follow-up Signal communication and calculation processing.
  • the first disadvantage of the prior art is high power consumption.
  • the execution of a large-scale neural network algorithm on a digital circuit requires a very large number of operations, resulting in very high power consumption.
  • applications such as robots can often only carry a limited amount of energy, and the power consumption of each part is required to be as low as possible. There is a contradiction between the two.
  • the second disadvantage of the prior art is the high complexity.
  • the source of the complexity is the complexity of the calculation operation.
  • the information of the calculation process needs to be processed repeatedly in the memory and the calculation unit to obtain the final result; the second is that the processing requires the environment
  • the analog information is collected and converted into digital signals, which are processed by calculation and then act on the output system, which has an impact on the environment. This conversion process also introduces redundant steps.
  • the high complexity of the technology limits the room for further improvement of its performance. Moreover, the high complexity also brings about the problem of too long system response time.
  • the present invention provides a robot control system based on a memristive cross-array, which solves the problems of high power consumption, high complexity, and slow response in the prior art.
  • the invention also provides a robot control method based on the memristive cross array.
  • the robot control system based on the memristive cross-array provided by the present invention includes a sensor group, an input sensor signal modulator, a neuromorphic circuit, an output control signal modulator, an output device, an external supervision module, and training control
  • the neuromorphic circuit includes N groups of cascaded memristor circuits, and the memristor circuit includes a memristor cross array, a differential amplifier circuit, and a multiplexing switch, where N ⁇ 1, specifically:
  • the sensor group is used to collect relevant information in the environment and convert the corresponding physical information in the environment into electrical quantities;
  • Input sensor signal modulator used to amplify the electrical quantity transmitted by each sensor to a suitable range, and modulate the signal waveform, the number of sensors corresponds to the number of channels of the input signal one-to-one;
  • the differential amplifier circuit and the multiplexing switch are both connected to the memristor cross array. After passing through the memristor cross array, the input signal vector is multiplied by the weight matrix stored in the memristor cross array, and then One or more analog output signals are obtained through a differential amplifier circuit, the multiplexing switch is used to perform a weight update operation on the memristor during the training process, and the other end of the multiplexing switch is connected to the training controller;
  • An output control signal modulator which is used to connect the output terminal of the neuromorphic circuit to modulate the analog output signal into a signal that matches and can drive the output device;
  • the output device is used to receive the modulated signal and drive the robot to produce the corresponding response
  • the external supervision module is used to judge whether the route of the robot is correct according to the response made by the robot during the training phase, and to control the positive and negative of the pulse entering the memristor;
  • the training controller is used to control the flow of electrical stimulation signals into related specific memristives according to the signals from the current sensors and the output signals of the neuromorphic circuit in the training phase, combined with the results of the external supervision module, and cooperate with the multiplex switch to control the flow of electrical stimulation signals into the relevant specific memristives.
  • the weight of the memristor is replaced.
  • the neuromorphic circuit has a neural network structure on the circuit structure, and the conductance value of the memristor in the neuromorphic circuit is used as the weight value of the synapse in the neural network.
  • the memristor cross array is composed of word lines and bit lines interleaved horizontally and longitudinally, and memristors at the cross points.
  • the memristor cross array allows the addition of other non-memristors while maintaining its circuit network structure. Electronic components do not violate the spirit of the present invention.
  • the system also includes a power supply module, which is responsible for all other modules to provide energy; a global clock module, which is used to coordinate other modules to work in an orderly manner.
  • the system also includes:
  • the pulse generator generates electrical excitation for updating the weight of the memristor, which is connected to the training controller.
  • the neuromorphic circuit also includes a non-linear element connected to the memristor array.
  • the present invention provides a robot control method based on memristive cross array, the method includes the following steps:
  • the sensor group collects relevant information in the environment and converts the corresponding physical quantities into electrical quantities as input signals;
  • the external supervision module supervises the behavior of the robot, and iterates steps (5) and (6) until the behavior mode of the robot is changed to the expected behavior mode;
  • the external supervision module determines whether the robot behavior is correct or not, and sends it to the training controller;
  • the training controller controls the flow of electrical stimulation signals into the relevant specific memristors, so that the weights of the memristors are replaced, and the result is to change the weight values of the neuromorphic circuit, which is expressed as if the behavior is correct, strengthen the greater weight. Decrease the smaller weight. If the behavior is wrong, the smaller weight will be strengthened and the larger weight will be weakened.
  • the neuromorphic circuit has a neural network structure on the circuit structure, and the conductance value of the memristor in the neuromorphic circuit is used as the weight value of the synapse in the neural network.
  • the memristor cross array is composed of word lines and bit lines interleaved horizontally and longitudinally, and memristors at the cross points.
  • the memristor cross array allows the addition of other non-memristors while maintaining its circuit network structure. Electronic components do not violate the spirit of the present invention.
  • the method also includes the power supply module providing energy to other modules; the global clock module coordinates the orderly cooperative work of various other modules.
  • the present invention has significant advantages: 1.
  • the present invention intends to use the memristor neuromorphic circuit to reduce the power consumption of the existing control system, improve the endurance of such intelligent robots, and perform The ability of larger-scale neural networks; 2.
  • Directly using memristors to simulate the weights in biological neural networks can effectively simplify the structure of the system, reduce complexity, and improve response speed.
  • Figure 1 is a block diagram of the information flow of the system of the present invention.
  • FIG. 2 is a schematic diagram of the structure and connection relationship of each circuit module included in the system of the present invention, in which an isolated module indicates that it has an effect on all other modules;
  • FIG. 3 is an electron microscope photograph of a memristor array with a neural network structure in an autopilot tracking robot based on a memristive neuromorphic circuit according to an embodiment of the present invention
  • FIG. 4 is a comparison diagram between the structure of the memristor cross array and the structure of a layer of fully connected neural network as an example in the embodiment of the present invention
  • Figure 5 is a photo of a prototype of an autonomous driving tracking robot based on memristive neuromorphic circuits in an embodiment of the present invention
  • Fig. 6 shows the weight transformation of the memristive synapse when the autopilot tracking robot prototype based on the memristive neuromorphic circuit in the implementation of the present invention performs a learning task
  • FIG. 7 is a comparison diagram of tracking task execution before and after learning and training of an autonomous driving tracking robot prototype based on a memristive neuromorphic circuit in an embodiment of the present invention
  • Fig. 8 is a characteristic diagram of the response speed of the memristive neuromorphic circuit in the embodiment of the present invention.
  • the robot control system based on the memristive cross-array provided by the present invention includes a sensor group, an input sensor signal modulator, a neuromorphic circuit, an output control signal modulator, an output device, an external supervision module, and a training controller.
  • the circuit includes N groups of cascaded memristor circuits, the memristor circuit includes a memristor cross array, a differential amplifier circuit, and a multiplexing switch, where N ⁇ 1, specifically:
  • Sensor group various types of sensors such as image, sound, temperature, pressure, etc. complexly detect relevant information in the environment, and convert the corresponding physical quantities in the environment into electrical quantities; the sensors in this system can be any type of any number of sensors, optical , Electricity, pressure, temperature, sound sensor.
  • Input sensor signal modulator used to amplify the electrical quantity transmitted by each sensor to a suitable range, and modulate the signal waveform, the number of sensors corresponds to the number of channels of the input signal one-to-one;
  • the differential amplifier circuit and the multiplexing switch are both connected to the memristor cross array. After passing through the memristor cross array, the input signal vector is multiplied by the weight matrix stored in the memristor cross array, and then One or more analog output signals are obtained through a differential amplifier circuit, the multiplexing switch is used to perform a weight update operation on the memristor during the training process, and the other end of the multiplexing switch is connected to the training controller;
  • the core of the neuromorphic circuit is composed of a memristor cross array, which is composed of word lines and bit lines that are staggered horizontally and vertically, and memristors at the crossing points.
  • a memristor cross array which is composed of word lines and bit lines that are staggered horizontally and vertically, and memristors at the crossing points.
  • other non-memristor electronic components such as transistors, diodes, etc., are allowed to be added, which does not violate the spirit of the present invention.
  • the neuromorphic circuit has a neural network structure in the circuit structure, and the conductance value of the memristor in the neuromorphic circuit is used as the weight value of the synapse in the neural network.
  • the memristive device here refers to a device whose resistance can be changed according to electrical stimulation.
  • Memristors of any material can be used in the present invention without departing from the spirit of the present invention and are within the protection scope of this patent.
  • neuromorphic circuit In order to realize the function of simulating neural synaptic network, neuromorphic circuit also needs some peripheral circuit modules to assist the work of memristor array.
  • peripheral circuit modules include signal differentiators, non-linear modules, and multiplexer switches.
  • the signal differentiator is used for differential processing of the two memristor signals, so that negative weights can be introduced into the memristive neural network.
  • Non-linear components can be added to introduce or enhance non-linearity in the multilayer network.
  • the neuromorphic circuit also needs a multi-channel check switch module, which is mainly used to update the weight of the memristor during the training process.
  • the size of the neural network and the number of layers are defined by the structure of the neuromorphic circuit.
  • the output control signal modulator is connected to the output terminal of the neuromorphic circuit, and modulates the analog output signal into a signal that matches and can drive the output device, such as pulse width signal, amplitude signal, frequency signal, etc.;
  • the output device receives the modulated signal and drives the robot to react accordingly. It may mainly include various types of motors, light emitting, sound, heating devices, and so on. Similarly, the types of output devices here are determined by actual application requirements, and different types of output devices do not violate the spirit of the present invention.
  • the external supervision module and training controller are used for supervised learning.
  • the external supervision module is used to judge whether the route of the robot is correct according to the response made by the robot during the training phase, and to control the positive and negative of the pulse entering the memristor;
  • the training controller is used to control the flow of electrical stimulation signals into related specific memristives according to the signals from the current sensors and the output signals of the neuromorphic circuit in the training phase, combined with the results of the external supervision module, and cooperate with the multiplex switch to control the flow of electrical stimulation signals into the relevant specific memristives.
  • the weight of the memristor is replaced, and the number of iterations is preset to complete the supervised learning of the system.
  • the behavior of the robot is that the robot can have different input-output mapping relationships, which ultimately enables the robot to master and optimize new skills under supervised learning.
  • the system also includes: a pulse generating device that generates electrical excitation for updating the weight of the memristor, which is connected to the training controller.
  • the system also includes a power supply module, which is responsible for all other modules to provide energy; a global clock module, which is used to coordinate other modules to work in an orderly manner.
  • This invention provides a new method for intelligent robot technology.
  • the operation of the biological neural network is directly simulated on the hardware circuit.
  • a memristive neuromorphic circuit is used as the control core of the robot, and an analog circuit based on a memristor with a neural network structure is used to generate the input-output mapping relationship, so that the robot can produce different actions to different environments.
  • neural networks have the ability to approximate arbitrary functions, and neural networks constructed on hardware by memristive neuromorphic circuits also have the same capabilities.
  • the memristive morphological neural circuit can adjust its internal connection parameters through the learning and training process, thereby approaching the ideal input-output mapping relationship.
  • the behavior of the robot can be expressed as that after the robot is in the process of supervised learning, it can learn the content taught in the supervised learning process, make expected responses in different environments, and can generalize to scenes that have not been learned before. , Make a reasonable and correct response.
  • the robot can be made with a simpler structure, achieving lower power consumption and faster response speed. So as to pave the way for the realization of robots with larger-scale neural networks.
  • the memristor with 40nm target/80nm tantalum/10nm tantalum oxide/40nm target structure is prepared by ultraviolet mask lithography technology, electron beam evaporation, magnetron sputtering, atomic layer deposition and other micro-processing technologies, as shown in the figure 3 is shown in the scanning electron microscope photo.
  • This kind of memristor has good adjustable properties, and can continuously change the resistance between 100 ohms and 4000 ohms after being excited by electrical pulses. Then use this memristor to build a cross-circuit structure, which is equivalent to a layer of fully connected neural network structure.
  • the picture on the left is a 3-by-3 memristor crossed array and the picture on the right is a layer of 3-by-3 fully connected neural network with the same topology, treating the memristor as a synapse, Regarding the electrode end point as the neuron end point, these two structures are completely the same.
  • a neuromorphic circuit is used as the one-layer structure of the neural network. It is also possible to consider using multiple layers of circuits as the multilayer of the neural network for experiments. Get learning results.
  • the robot car shown in Figure 5 was built as a prototype based on the memristor neuromorphic car.
  • the modules shown in FIG. 5 are integrated and assembled by means of printed circuit boards. The details are described below.
  • Two light intensity sensors are installed on the front of the chassis of the robot car. Based on the principle of a photoresistor, the sensor can detect the brightness of the road in front of the chassis. The detected light intensity information is converted into a voltage signal. When the light changes from strong to weak, the voltage changes from high to low. The voltage signal is reduced to within 0.2V and then input into the neuromorphic circuit for processing.
  • the memristor array is connected to the printed circuit board by wire bonding.
  • the output signal After passing through the memristor cross-array circuit, the output signal is the product of the input signal and the weight matrix stored in the memristor array, and then passes through the differential amplifier circuit to obtain an analog output signal.
  • the analog output signal is modulated into a pulse width modulation signal under the action of the output control modulator to further control the steering of the front wheel steering gear.
  • the robot in the working mode of the robot, the robot will process the signals detected in real time and respond to the direction of rotation of the front wheels. It is worth noting that the control system in the assembly and construction of the robot is completely based on memristive neuromorphic circuits without any digital processor, which shows that this example strictly complies with the design of the present invention.
  • the robot can be set to learning mode to train the robot to perform specific tasks.
  • the onboard pulse signal generator outputs a pulse signal, and according to the set weight update method, the pulse signal is input to the memristor through the multi-channel multiplexer to realize the weight update.
  • the robot car will be trained to perform tracking tasks.
  • the robot is placed in a special environment, and the external supervision module judges the correctness of its response according to the response of the robot, and controls the positive and negative of the pulse entering the memristor. Every time all memristors are traversed is called an iterative process.
  • the maximum weight in the iteration process and the memristor weight in each iteration are recorded, and the weight is normalized by the memristor weight/maximum weight; the 4 memristors in this example have experienced 37 After the second iteration process, the weight value has changed significantly.
  • the behavior of the robot car has also changed as shown in Figure 7. From the inability to perform the task shown on the left side of Figure 7 to the right side of Figure 7, the robot car can complete the tracking task completely, showing that the car as an intelligent robot has the ability to learn to perform specific tasks, which verifies the effectiveness of the present invention And feasibility.
  • the inventor tested the response speed of the memristive neuromorphic circuit used in the robot car, as shown in Figure 8. It can be seen that under the half-swing point as a measurement standard, the response speed of the circuit from input to output averages at 56ns, which is a hundredfold advantage compared to the usual embedded single-chip processors. This result shows that based on a simpler structure and arithmetic process, the memristive neuromorphic circuit of the present invention can achieve faster real-time response speed.
  • the present invention also provides a robot control method based on memristive cross array, which includes the following steps:
  • the sensor group collects relevant information in the environment and converts the corresponding physical quantities into electrical quantities as input signals;
  • the external supervision module supervises the behavior of the robot, and iterates steps (5) and (6) until the behavior mode of the robot is changed to the expected behavior mode;
  • the external supervision module determines whether the robot behavior is correct or not, and sends it to the training controller;
  • the training controller controls the electrical stimulation signal to flow into the relevant specific memristor, so that the weight of the memristor is replaced. The result is to change the weight value of the neuromorphic circuit. If the behavior is correct, strengthen the larger weight and weaken the smaller weight. If the behavior is wrong, the smaller weight will be strengthened and the larger weight will be weakened.
  • the neuromorphic circuit has a neural network structure on the circuit structure, and the conductance value of the memristor in the neuromorphic circuit is used as the weight value of the synapse in the neural network.
  • the interleaved memristor array is composed of word lines and bit lines interleaved horizontally and vertically, and memristors at intersections, and further includes:
  • the method also includes the power supply module providing energy to other modules; the global clock module coordinates the orderly cooperative work of various other modules.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Robotics (AREA)
  • Mechanical Engineering (AREA)
  • Automation & Control Theory (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Fuzzy Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Neurology (AREA)
  • Human Computer Interaction (AREA)
  • Feedback Control In General (AREA)
  • Manipulator (AREA)

Abstract

一种基于忆阻交叉阵列的机器人控制系统及方法，该系统包括传感器组、输入传感信号调制器、神经形态电路、输出控制信号调制器、输出装置、外部监督模块和训练控制器；机器人的控制由神经形态电路执行，其主要部分是具有全连接神经网络结构的忆阻器交叉阵列。神经形态电路中的差分放大电路和多路复用开关均连接在忆阻器交叉阵列上，输入信号向量与忆阻器交叉阵列内储存的权重矩阵的相乘，再经过差分放大电路得到一路或者多路模拟输出信号。该控制系统利用忆阻器神经形态电路来降低已有机器人控制系统执行神经网络运算时的功耗；直接用忆阻器来模拟生物神经网络中的权重，有效降低复杂度，提高了响应速度。

Description

一种基于忆阻交叉阵列的机器人控制系统及方法 技术领域
本发明涉及忆阻交叉阵列自动控制领域，具体涉及一种基于忆阻交叉阵列的机器人控制系统及方法。
背景技术
现有的智能机器人技术在硬件上是以数字处理器为核心的数字控制电路，在软件上是基于各种神经网络算法实现智能应用。作为这类以神经网络算法驱动的机器人，要求其处理时的功耗低，速度快；在目前的机器人等应用平台上，依靠已有的数字控制电路驱动神经网络算法给其硬件带来了严峻的挑战。
在这类智能机器人的软件层面上，包含了复杂的信息处理和大量的计算。其信息控制流按照以下描述进行。通过各类图像、温度、压力、声音等传感器，采集环境当中的物理量并将其转化为数字信号，通过数字通信传递到数字处理器中。在数字处理器中，按照神经网络算法对信息进行处理。在通用的神经网络算法模型中，由若干层结构类似的处理过程组成。每层网络中都包含着大量的权重参数，在每层网络中的数字信息和网络权重参数进行矩阵乘法运算以及非线性运算，得到该层的输出，进一步转化成为下一层的输入。在经历了神经网络算法后，输入信息转化为了计算结果，再依照输出结果调制控制信号，控制机器人的各种智能动作。为了使得机器人可以有适应和学习能力，该神经网络还需进行一定的反馈训练。最常见的方法就是误差反向传播和梯度下降方法，而这些训练方法涉及大量的参数调整和计算。
在这类机器人的硬件层面上，主要有负责传感信息，负载处理信息，和负责运动控制的部分组装而成。其中负责信息处理的核心主要设计数字控制器电路。其包含的集成数字芯片都是基于互补金属氧化物半导体技术，以晶体管作为基本元件制成的逻辑电路，其主要有通用处理器和内存等电路模块。在进行计算的时候，限制其性能的主要瓶颈之一是由于内存和处理器之间有限的信息交换速度，在进行训练阶段时尤为明显。对于传感部分，往往传感器是通过物理效应制成，元件收到环境中的物理量的影响，具有着不同的电学表现，这种响应体现在电学上是模拟信号。但是为了和数字处理系统匹配，传感器元件将环境中的模拟量转化为电学模拟量后，还需要配合模数转换器转换为数字信号，以通信的形式传递到后级的控制电路以供后续的信号通信和计算处理。
现有技术的缺点一是功耗高，在数字电路上执行规模较大的神经网络算法需要极大的运算次数，导致了非常高的功耗。而机器人这类应用往往只可以携带有限能源，要求其各部分功耗尽可能要低。两者之间产生矛盾。
现有技术的缺点二是复杂度高，其中复杂度的来源一是计算操作的复杂度，为计算过程信息需要在内存和计算单元中反复进行处理以得到最终结果；二是处理需要将环境中的模拟信息采集转化为数字信号并进行运算处理后作用于输出系统，对环境产生影响。这一转化过程也引入了多余的步骤。技术的高复杂度限制了其性能进一步提高的空间。而且，高复杂度同时也带来了系统的响应时间过长的问题。
发明内容
发明目的：为了克服现有技术的不足，本发明提供一种基于忆阻交叉阵列的机器人控制系统，该系统解决了现有技术中存在的功耗高，复杂度高，响应慢的问题，本发明还提供一种基于忆阻交叉阵列的机器人控制方法。
技术方案：一方面，本发明提供的基于忆阻交叉阵列的机器人控制系统，包括传感器组、输入传感信号调制器、神经形态电路、输出控制信号调制器、输出装置、外部监督模块和训练控制器，所述神经形态电路包括N组级联的忆阻器电路，所述忆阻器电路包括忆阻器交叉阵列、差分放大电路以及多路复用开关，其中，N≥1，具体的：
传感器组，用于采集环境中的相关信息，并将环境中的相应物理信息量转化为电学量；
输入传感信号调制器，用于放大各个传感器传递的电学量到适合范围内，并对信号波形进行调制，传感器的数量和输入信号的通道数一一对应；
所述差分放大电路和多路复用开关均连接在所述忆阻器交叉阵列上，经过忆阻器交叉阵列之后，输入信号向量与忆阻器交叉阵列内储存的权重矩阵的相乘，再经过差分放大电路得到一路或者多路模拟输出信号，所述多路复用开关，用于在训练过程中用来对忆阻器进行权重更新操作，其另一端与所述训练控制器相连；
输出控制信号调制器，用于连接所述神经形态电路的输出端，将模拟输出信号调制为匹配并可以驱动输出装置的信号；
输出装置，用于接受到调制好的信号，驱动机器人产生相应的反应；
外部监督模块，用于在训练阶段根据机器人做出的响应来判断其路线是否正确，控制进入忆阻器的脉冲的正负性；
训练控制器，用于在训练阶段根据当前传感器给出的信号和神经形态电路的输出信号，并结合外部监督模块的结果，配合所述多路复用开关控制电学刺激信号流入相关的特定忆阻器，使得忆阻器的权重产生更替。
进一步地，包括：
所述神经形态电路在电路结构上具有神经网络的结构，用所述神经形态电路中忆阻器的电导值作为神经网络中突触的权重值。
进一步地，包括：
所述忆阻器交叉阵列由横纵交错的字线和位线以及交叉点处的忆阻器构成，忆阻器交叉阵列在保持其电路网络结构的前提下，允许加入其他非忆阻器的电子元件，不违背本发明的精神。
进一步地，包括：
该系统还包括供电模块，其用于对其他所有模块负责以提供能源；全局时钟模块，用于协调其他各个不同模块有序协同工作。
进一步地，包括：
该系统还包括：
脉冲产生装置，产生用于忆阻器权重更新的电学激励，其与训练控制器相连。
进一步地，包括：
所述神经形态电路还包括与所述忆阻器阵列相连的非线性元件。
另一方面，本发明提供一种基于忆阻交叉阵列的机器人控制方法，该方法包括以下步骤：
(1)传感器组采集环境中的相关信息后将相应的物理量转换为电学量，作为输入信号；
(2)将输入信号输入到输入传感信号调制器进行调制，得到调制后的输入信号；
(3)将调制后的输入信号输入到神经形态电路的忆阻器交叉阵列中，经过忆阻器交叉阵列之后，输出的信号为输入信号与忆阻器交叉阵列内储存的权重矩阵的乘积，随后再经过差分放大电路得到一路模拟输出信号，输出装置接受到所述模拟输出信号，驱动机器人产生相应的反应；
(4)外部监督模块监督机器人的行为，并迭代步骤(5)和(6)，直到将机器人的行为模式改变为预期行为模式；
(5)外部监督模块判定机器人行为正确与否，并发送给训练控制器；
(6)训练控制器控制电学刺激信号流入到相关的特定忆阻器，使得忆阻器的权重产生更替，其结果是改变神经形态电路的权重值，表现为如果行为正确，加强较大权重，减弱较小权重。如果行为错误，则会加强较小权重，减弱较大权重。
进一步地，包括：
所述神经形态电路在电路结构上具有神经网络的结构，用所述神经形态电路中忆阻器的电导值作为神经网络中突触的权重值。
进一步地，包括：
所述忆阻器交叉阵列由横纵交错的字线和位线以及交叉点处的忆阻器构成，忆阻器交叉阵列在保持其电路网络结构的前提下，允许加入其他非忆阻器的电子元件，不违背本发明的精神。
进一步地，包括：
所述该方法还包括供电模块对其他模块提供能源；全局时钟模块协调其他各个不同模块有序协同工作。
有益效果：本发明与现有技术相比，其显著优点是：1、本发明拟利用忆阻器神经形态电路来降低已有控制系统的功耗，提高此类智能机器人的续航能力，以及执行更大规模神经网络的能力；2、直接用忆阻器来模拟生物神经网络中的权重，可以有效简化系统的结构，降低复杂度，提高了响应速度。
附图说明
图1是本发明所述的系统的信息流框图；
图2是本发明所述的系统中包含各个电路模块的结构与连接关系示意图，其中孤立的模块表示对其他所有模块都起作用；
图3是本发明一实施例所述的基于忆阻神经形态电路的自动驾驶循迹机器人中具有神经网络结构的忆阻器阵列的电子显微镜照片；
图4是本发明实施实例中举例的忆阻器交叉阵列的结构和一层全连接神经网络结构的对比图；
图5是本发明实施实例基于忆阻神经形态电路的自动驾驶循迹机器人原型机的照片；
图6是本发明实施实例基于忆阻神经形态电路的自动驾驶循迹机器人原型机在执行 学习任务时忆阻突触的权重变换；
图7是本发明实施实例基于忆阻神经形态电路的自动驾驶循迹机器人原型机在学习训练前后的循迹任务执行对比图；
图8是本发明实施实例中的忆阻神经形态电路的响应速度表征图。
具体实施方式
本发明提供的基于忆阻交叉阵列的机器人控制系统，包括传感器组、输入传感信号调制器、神经形态电路、输出控制信号调制器、输出装置、外部监督模块和训练控制器，所述神经形态电路包括N组级联的忆阻器电路，所述忆阻器电路包括忆阻器交叉阵列、差分放大电路以及多路复用开关，其中N≥1，具体的：
传感器组，图像、声音、温度、压力等各类传感器复杂探测环境中的相关信息，并将环境中的相应的物理量转化为电学量；本系统中的传感器可以是任意种类任意数量的传感器，光学、电学、压力、温度、声音传感器。
输入传感信号调制器，用于放大各个传感器传递的电学量到适合范围内，并对信号波形进行调制，传感器的数量和输入信号的通道数一一对应；
所述差分放大电路和多路复用开关均连接在所述忆阻器交叉阵列上，经过忆阻器交叉阵列之后，输入信号向量与忆阻器交叉阵列内储存的权重矩阵的相乘，再经过差分放大电路得到一路或者多路模拟输出信号，所述多路复用开关，用于在训练过程中用来对忆阻器进行权重更新操作，其另一端与所述训练控制器相连；
具体的，神经形态电路核心由忆阻器交叉阵列构成，忆阻器交叉阵列是横纵交错的字线和位线以及交叉点处的忆阻器构成。忆阻器交叉阵列在保持其电路网络结构的前提下，允许加入其他非忆阻器的电子元件，如晶体管、二极管等，不违背本发明的精神。神经形态电路在电路结构上具有神经网络的结构，用所述神经形态电路中忆阻器的电导值作为神经网络中突触的权重值。
同样，这里的忆阻器件指代电阻可以根据电学刺激而改变的器件，任意材料的忆阻器都可以在本发明中使用，均不违背本发明的精神，在本专利的保护范围之内。
神经形态电路为了实现模拟神经突触网络的功能，还需要一些外围电路模块辅助忆阻器阵列的工作。这些外围电路模块包含信号差分器、非线性模块以及多路复选开关。其中信号差分器用于将两路忆阻器信号做差分处理，这样可以在忆阻神经网络中引入负值权重。可添加非线性元件，在多层网络中引入或增强非线性。
另外，神经形态电路中还需要多路复选开关模块，主要在训练过程中用来对忆阻器进行权重更新操作。神经网络的规模大小，层数由神经形态电路的结构定义。
输出控制信号调制器连接神经形态电路的输出端，将模拟输出信号调制为匹配并可以驱动输出装置的信号，比如脉宽信号、幅值信号、频率信号等等；
输出装置接受到调制好的信号，驱动机器人产生相应的反应。可能主要包含的有各类马达电机，发光、发声、发热装置等等。同样的，这里的输出装置的种类由实际的应用需求决定，不同种类的输出装置均不违背本发明的精神。
外部监督模块和训练控制器用于监督学习。
外部监督模块，用于在训练阶段根据机器人做出的响应来判断其路线是否正确，控制进入忆阻器的脉冲的正负性；
训练控制器，用于在训练阶段根据当前传感器给出的信号和神经形态电路的输出信号，并结合外部监督模块的结果，配合所述多路复用开关控制电学刺激信号流入相关的特定忆阻器，使得忆阻器的权重产生更替，预设迭代次数，完成系统的监督学习。表现在机器人的行为上就是机器人可以拥有不同的输入-输出映射关系，最终使得机器人在有监督学习下掌握和优化新的技能。
进一步地，包括：
该系统还包括：脉冲产生装置，产生用于忆阻器权重更新的电学激励，其与训练控制器相连。
该系统还包括供电模块，其用于对其他所有模块负责以提供能源；全局时钟模块，用于协调其他各个不同模块有序协同工作。
该发明为智能机器人技术提供了一个新的方法。通过忆阻电路的神经网络结构的连接，和忆阻器对模拟信号的调制作用，直接在硬件电路上模拟生物神经网络的运行。本发明中利用了一个忆阻神经形态电路作为机器人的控制核心，利用基于忆阻器的，具有神经网络结构的模拟电路来产生输入输出映射关系，使得机器人对不同环境产生不同的动作。理论上神经网络具有逼近任意函数的能力，而忆阻神经形态电路在硬件上构建的神经网络也具有相同的能力。在本发明中，忆阻形态神经电路可以通过学习训练过程来调整其内部的连接参数，从而逼近理想的输入输出映射关系。在机器人的行为上即可以表现为机器人在监督学习的过程后，可以学习到监督学习过程中教授的内容，在不用的环境下做出期望的反应，并且可以泛化到之前没有学习过的场景，做出合理正确的反应。
利用这种基于忆阻神经形态电路的新方法，制作出的机器人可以以更简单的结构，实现更低功耗和更快的反应速度。从而为实现具有规模更大的神经网络的机器人开辟道路。
以下为实施本发明提供的控制系统下制作的巡径机器人的一个具体实施例：
一个基于忆阻神经形态电路机器人小车的原型机，并且展示了该机器人可以通过学习过程具备循迹能力，不仅验证了该发明的可行性，同时也得到了忆阻神经形态电路具有极快的响应速度(约56ns)的实验结论。具体实施过程如下：
首先通过紫外掩模光刻技术，电子束蒸镀，磁控溅射，原子层沉积等微加工技术，制备出具有40nm靶/80nm钽/10nm氧化钽/40nm靶结构的忆阻器，如图3中的扫描电子显微镜照片所示。该种忆阻器具有良好的可调性质，受到电学脉冲激励后可以在100欧姆-4000欧姆之间连续改变阻值。然后利用这种忆阻器构建起交叉电路结构，这种结构等价于一层全连接神经网络的结构。如图4所示，左边的图是一个3乘3的忆阻器交叉阵列和右边的图是一层3乘3的全连接神经网络有相同的拓扑结构，将忆阻器视为突触，将电极端点视为神经元端点，这两种结构完全相同，本实施例采用一个神经形态电路作为神经网络的一层结构，也可考虑采用多个层电路作为神经网络的多层进行实验，得到学习结果。
依照图2所示的结构，搭建了图5所示的机器人小车作为基于忆阻器神经形态小车的原型机。
在本实施实例中，图5所示的各模块依靠印刷电路板的方式进行集成装配。具体在如下描述。在机器人小车的底盘前部安装有两个光强传感器，基于光敏电阻的原理，该传感器可以探测到底盘前部道路的明暗。探测到的光强信息被转化为了电压信号，光线由强到弱变化时，电压由高到低。该电压信号被缩小到0.2V以内后输入了神经形态电路中进行处理。忆阻器阵列通过焊线的方式引接在印刷电路板上。经过忆阻器交叉阵列电路之后，输出的信号为输入信号与忆阻器阵列内储存的权重矩阵的乘积，随后再经过差分放大电路得到一路模拟输出信号。该模拟输出信号在输出控制调制器的作用下，调制成为脉宽调制信号，进一步控制前轮舵机的转向。在该实例中，在该机器人的工作模式下，机器人会将实时探测到的信号进行处理，并以前轮的转动方向作为响应。值得注意的是，该机器人的组装搭建中的控制系统完全基于忆阻神经形态电路，没有采用任何数字处理器，这体现了该实例严格遵守本发明中的设计。
在该实例中，可以将机器人设置到学习模式，以训练改机器人执行特定任务。在该模式下，板载的脉冲信号发生器输出脉冲信号，按照设定的权值更新方法，通过多通道复用器将脉冲信号输入忆阻器以实现权值更新。在该实施例中，该机器人小车将被训练执行循迹任务。该机器人被放置在特殊环境中，外部监督模块根据机器人做出的响应来判断其响应的正确与否，控制进入忆阻器的脉冲的正负性。每遍历过所有的忆阻器称为一次迭代过程。如图6所示，记录迭代过程中的最大权重以及每次迭代的忆阻器权重，采用忆阻器权重/最大权重对权重进行归一化；该实例中的4个忆阻器经历了37次迭代过程后权重值发生了明显改变。而机器人小车的行为也发生了图7所示的变化。从图7左边所示的无法执行任务到图7右边所示，该机器人小车可以完整的完成循迹任务，展示了该小车作为智能机器人具有学习执行特定任务的能力，验证了本发明的有效性和可行性。
最后，发明人测试了该机器人小车用到的忆阻神经形态电路的响应速度，如图8所示。可以看到，以半摆幅点作为衡量标准下，该电路从输入到输出的响应速度平均在56ns的水平，对比通常的嵌入式单片处理器这已经有百倍的优势。该结果说明，基于更简单的结构和运算过程，本发明中的忆阻神经形态电路可以实现更快的实时响应速度。
本发明还提供一种基于忆阻交叉阵列的机器人控制方法，该方法包括以下步骤：
(1)传感器组采集环境中的相关信息后将相应的物理量转换为电学量，作为输入信号；
(2)将输入信号输入到输入传感信号调制器进行调制，得到调制后的输入信号；
(3)将调制后的输入信号输入到神经形态电路的忆阻器交叉阵列中，经过忆阻器交叉阵列之后，输入信号向量与忆阻器交叉阵列内储存的权重矩阵的相乘，再经过差分放大电路得到一路或者多路模拟输出信号，驱动机器人产生相应的反应；
(4)外部监督模块监督机器人的行为，并迭代步骤(5)和(6)，直到将机器人的行为模式改变为预期行为模式；
(5)外部监督模块判定机器人行为正确与否，并发送给训练控制器；
(6)训练控制器控制电学刺激信号流入到相关的特定忆阻器，使得忆阻器的权重产生更替。其结果是改变神经形态电路的权重值，表现为如果行为正确，加强较大权重，减弱较小权重。如果行为错误，则会加强较小权重，减弱较大权重。
进一步地，包括：
所述神经形态电路在电路结构上具有神经网络的结构，用所述神经形态电路中忆阻器的电导值作为神经网络中突触的权重值。
进一步地，包括：
所述忆阻器交叉阵列由横纵交错的字线和位线以及交叉点处的忆阻器构成，进一步地，包括：
所述该方法还包括供电模块对其他模块提供能源；全局时钟模块协调其他各个不同模块有序协同工作。

Claims (10)

  1. 一种基于忆阻交叉阵列的机器人控制系统，其特征在于，包括传感器组、输入传感信号调制器、神经形态电路、输出控制信号调制器、输出装置、外部监督模块和训练控制器，所述神经形态电路包括N组级联的忆阻器电路，所述忆阻器电路包括忆阻器交叉阵列、差分放大电路以及多路复用开关，其中，N≥1，具体的：
    传感器组，用于采集环境中的相关信息，并将环境中的相应物理信息量转化为电学量；
    输入传感信号调制器，用于放大各个传感器传递的电学量到适合范围内，并对信号波形进行调制，传感器的数量和输入信号的通道数一一对应；
    所述差分放大电路和多路复用开关均连接在所述忆阻器交叉阵列上，经过忆阻器交叉阵列之后，输入信号向量与忆阻器交叉阵列内储存的权重矩阵的相乘，再经过差分放大电路得到一路或者多路模拟输出信号；所述多路复用开关，用于在训练过程中对忆阻器进行权重更新操作，其另一端与所述训练控制器相连；
    输出控制信号调制器，用于连接所述神经形态电路的输出端，将模拟输出信号调制为匹配并可以驱动输出装置的信号；
    输出装置，用于接受到调制好的信号，驱动机器人产生相应的反应；
    外部监督模块，用于在训练阶段根据机器人做出的响应来判断其路线是否正确，控制进入忆阻器的脉冲的正负性；
    训练控制器，用于在训练阶段根据当前传感器给出的信号和神经形态电路的输出信号，并结合外部监督模块的结果，配合所述多路复用开关控制电学刺激信号流入相关的特定忆阻器，使得忆阻器的权重产生更替。
  2. 根据权利要求1所述的基于忆阻交叉阵列的机器人控制系统，其特征在于，所述神经形态电路在电路结构上具有神经网络的结构，用所述神经形态电路中忆阻器的电导值作为神经网络中突触的权重值。
  3. 根据权利要求1所述的基于忆阻交叉阵列的机器人控制系统，其特征在于，所述忆阻器交叉阵列由横纵交错的字线和位线以及交叉点处的忆阻器构成，所述忆阻器交叉阵列在保持其电路网络结构的前提下，允许加入其他非忆阻器的电子元件。
  4. 根据权利要求1所述的基于忆阻交叉阵列的机器人控制系统，其特征在于，该系统还包括供电模块，其用于对其他所有模块负责以提供能源；全局时钟模块，用于协调其他各个不同模块有序协同工作。
  5. 根据权利要求1所述的基于忆阻交叉阵列的机器人控制系统，其特征在于，该系统还包括：
    脉冲产生装置，产生用于忆阻器权重更新的电学激励，其与训练控制器相连。
  6. 根据权利要求1所述的基于忆阻交叉阵列的机器人控制系统，其特征在于，所述神经形态电路还包括与所述忆阻器阵列相连的非线性元件。
  7. 一种基于忆阻交叉阵列的机器人控制方法，其特征在于，该方法包括以下步骤：
    (1)传感器组采集环境中的相关信息后将相应的物理量转换为电学量，作为输入信号；
    (2)将输入信号输入到输入传感信号调制器进行调制，得到调制后的输入信号；
    (3)将调制后的输入信号输入到神经形态电路的忆阻器交叉阵列中，经过忆阻器交叉阵列之后，输入信号向量与忆阻器交叉阵列内储存的权重矩阵的相乘，再经过差分放大电路得到一路或者多路模拟输出信号，输出装置接受到所述模拟输出信号，驱动机器人产生相应的反应；
    (4)外部监督模块监督机器人的行为，并迭代步骤(5)和(6)，直到将机器人的行为模式改变为预期行为模式；
    (5)外部监督模块判定机器人行为正确与否，并发送给训练控制器；
    (6)训练控制器控制电学刺激信号流入到相关的特定忆阻器，使得忆阻器的权重产生更替，其结果是改变神经形态电路的权重值，表现为：如果行为正确，加强较大权重，减弱较小权重；如果行为错误，则会加强较小权重，减弱较大权重。
  8. 根据权利要求7所述的基于忆阻交叉阵列的机器人控制方法，其特征在于，所述神经形态电路在电路结构上具有神经网络的结构，用所述神经形态电路中忆阻器的电导值作为神经网络中突触的权重值。
  9. 根据权利要求7所述的基于忆阻交叉阵列的机器人控制方法，其特征在于，所述忆阻器交叉阵列由横纵交错的字线和位线以及交叉点处的忆阻器构成，忆阻器交叉阵列在保持其电路网络结构的前提下，允许加入其他非忆阻器的电子元件。
  10. 根据权利要求7所述的基于忆阻交叉阵列的机器人控制方法，其特征在于，所述该方法还包括供电模块对其他模块提供能源；全局时钟模块协调其他各个不同模块有序协同工作。
PCT/CN2019/114958 2019-10-18 2019-11-01 一种基于忆阻交叉阵列的机器人控制系统及方法 WO2021072817A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/763,495 US20220331952A1 (en) 2019-10-18 2019-11-01 System and method for robot control based on memristive crossbar array

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910993530.3A CN110842915B (zh) 2019-10-18 2019-10-18 一种基于忆阻交叉阵列的机器人控制系统及方法
CN201910993530.3 2019-10-18

Publications (1)

Publication Number Publication Date
WO2021072817A1 true WO2021072817A1 (zh) 2021-04-22

Family

ID=69597470

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/114958 WO2021072817A1 (zh) 2019-10-18 2019-11-01 一种基于忆阻交叉阵列的机器人控制系统及方法

Country Status (3)

Country Link
US (1) US20220331952A1 (zh)
CN (1) CN110842915B (zh)
WO (1) WO2021072817A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115019856A (zh) * 2022-08-09 2022-09-06 之江实验室 一种基于rram多值存储的存内计算方法与系统
WO2022225465A1 (en) * 2021-04-23 2022-10-27 Nanyang Technological University Systems for near-sensor analogue computing for ultrafast responsive tactile sensing

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111460365B (zh) * 2020-03-10 2021-12-03 华中科技大学 一种基于忆阻线性神经网络的方程组求解器及其操作方法
CN113311702B (zh) * 2021-05-06 2022-06-21 清华大学 一种基于Master-Slave神经元的人工神经网络控制器
CN113296952B (zh) * 2021-06-01 2022-03-18 南京大学 一种采用模拟传感器高阶微分进行边缘计算的系统及方法
CN113506850B (zh) * 2021-06-01 2024-02-09 北京航空航天大学合肥创新研究院(北京航空航天大学合肥研究生院) 一种具有非单调变化电阻态的忆阻器在碰撞预测中的应用
CN113465635A (zh) * 2021-06-10 2021-10-01 电子科技大学 一种传感器与神经网络计算单元信号转换的装置
CN114279491A (zh) * 2021-11-23 2022-04-05 电子科技大学 一种基于忆阻交叉阵列的传感器信号注意力权重分配方法
CN117648960B (zh) * 2024-01-30 2024-04-19 中国人民解放军国防科技大学 基于忆阻器的脉冲神经网络在线训练电路及方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090163826A1 (en) * 2006-04-03 2009-06-25 Blaise Laurent Mouttet Memristor crossbar neural interface
CN102354128A (zh) * 2011-06-02 2012-02-15 北京大学 一种用于机器人情感模拟的电路及其控制方法
CN102723112A (zh) * 2012-06-08 2012-10-10 西南大学 一种基于忆阻交叉阵列的q学习系统
US20170011290A1 (en) * 2015-07-06 2017-01-12 University Of Dayton On-chip training of memristor crossbar neuromorphic processing systems
CN108009640A (zh) * 2017-12-25 2018-05-08 清华大学 基于忆阻器的神经网络的训练装置及其训练方法
CN109800870A (zh) * 2019-01-10 2019-05-24 华中科技大学 一种基于忆阻器的神经网络在线学习系统

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104809498B (zh) * 2014-01-24 2018-02-13 清华大学 一种基于神经形态电路的类脑协处理器
CN105306192A (zh) * 2015-10-14 2016-02-03 常州大学 一种耦合一阶广义忆阻器实现的四阶忆阻Colpitts混沌信号发生器

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090163826A1 (en) * 2006-04-03 2009-06-25 Blaise Laurent Mouttet Memristor crossbar neural interface
CN102354128A (zh) * 2011-06-02 2012-02-15 北京大学 一种用于机器人情感模拟的电路及其控制方法
CN102723112A (zh) * 2012-06-08 2012-10-10 西南大学 一种基于忆阻交叉阵列的q学习系统
US20170011290A1 (en) * 2015-07-06 2017-01-12 University Of Dayton On-chip training of memristor crossbar neuromorphic processing systems
CN108009640A (zh) * 2017-12-25 2018-05-08 清华大学 基于忆阻器的神经网络的训练装置及其训练方法
CN109800870A (zh) * 2019-01-10 2019-05-24 华中科技大学 一种基于忆阻器的神经网络在线学习系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
GU BEN J.; HUANG XIN; OU AMBER; REMBACH ALAN; FOWLER CHRISTOPHER; AVULA PAVAN K.; HORTON ADAM; DOECKE JAMES D.; VILLEMAGNE VICTOR : "Innate phagocytosis by peripheral blood monocytes is altered in Alzheimer’s disease", ACTA NEUROPATHOLOGICA, SPRINGER VERLAG, BERLIN, DE, vol. 132, no. 3, 13 July 2016 (2016-07-13), BERLIN, DE, pages 377 - 389, XP036036918, ISSN: 0001-6322, DOI: 10.1007/s00401-016-1596-3 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022225465A1 (en) * 2021-04-23 2022-10-27 Nanyang Technological University Systems for near-sensor analogue computing for ultrafast responsive tactile sensing
CN115019856A (zh) * 2022-08-09 2022-09-06 之江实验室 一种基于rram多值存储的存内计算方法与系统
CN115019856B (zh) * 2022-08-09 2023-05-16 之江实验室 一种基于rram多值存储的存内计算方法与系统

Also Published As

Publication number Publication date
US20220331952A1 (en) 2022-10-20
CN110842915B (zh) 2021-11-23
CN110842915A (zh) 2020-02-28

Similar Documents

Publication Publication Date Title
WO2021072817A1 (zh) 一种基于忆阻交叉阵列的机器人控制系统及方法
US11861489B2 (en) Convolutional neural network on-chip learning system based on non-volatile memory
CN105224986B (zh) 基于忆阻器件的深度神经网络系统
Rajendran et al. Specifications of nanoscale devices and circuits for neuromorphic computational systems
CN106981567B (zh) 一种基于光电耦合忆阻器的人工突触器件及其调制方法
US8515885B2 (en) Neuromorphic and synaptronic spiking neural network with synaptic weights learned using simulation
JP6764473B2 (ja) 抵抗型処理ユニット
CN113272828A (zh) 弹性神经网络
EP3121767A1 (en) Methods and systems for implementing deep spiking neural networks
Stewart et al. Online few-shot gesture learning on a neuromorphic processor
Merrikh-Bayat et al. Memristive neuro-fuzzy system
Glatz et al. Adaptive motor control and learning in a spiking neural network realised on a mixed-signal neuromorphic processor
Dalton et al. Artificial neural networks
Chan et al. A neural network controller for switching power converters
Zhang et al. Neuromorphic circuit implementation of operant conditioning based on emotion generation and modulation
US6876989B2 (en) Back-propagation neural network with enhanced neuron characteristics
Hsu et al. Recurrent fuzzy-neural approach for nonlinear control using dynamic structure learning scheme
Wang et al. Unsupervising adaption neural-network control
JP2716796B2 (ja) 光コンピユータ
CN117151176A (zh) 用于神经网络学习的突触阵列、运算电路及操作方法
Yang et al. Neuromorphic electronics for robotic perception, navigation and control: A survey
US11250316B2 (en) Aggregate adjustments in a cross bar neural network
JPH04237388A (ja) ニューロプロセッサ
CA2898216C (en) Methods and systems for implementing deep spiking neural networks
Oubbati et al. Meta-learning for adaptive identification of non-linear dynamical systems

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19949401

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19949401

Country of ref document: EP

Kind code of ref document: A1