85 results about "Reservoir computing" patented technology

The invention relates to an AI processor and method based on storage and calculation integration, three-dimensional integration and operator separation. The AI processor comprises a storage layer; the calculation layer and the storage layer are stacked in the vertical direction through a three-dimensional integrated connection structure, and a three-dimensional integrated high-speed channel is generated and used for executing calculation tasks in a large language model; the standardized memory interface is used for being connected with an external main control chip; the calculation task comprises a pre-filling stage and a decoding stage; the scheduling module configures the main control chip to interact data with the storage layer through the standardized memory interface so as to execute the pre-filling stage; the scheduling module configures the computing layer to interact data with the storage layer through the three-dimensional integrated high-speed channel so as to execute the decoding stage, the pre-filling stage is processed by utilizing the large computing power advantage of a main control chip, and the decoding stage is processed by utilizing the three-dimensional stacked high-bandwidth advantage, so that accurate matching of the computing power and the bandwidth is realized; and the large model reasoning efficiency is obviously improved.

The invention belongs to the technical field of reservoir calculation, and relates to a multi-dimensional pulse reservoir calculation method and system based on a semiconductor laser. The multi-dimensional pulse reservoir calculation method comprises the following steps: encoding a signal to be processed to generate an original optical signal; injecting the original optical signal into a semiconductor laser, and collecting a spike pulse optical signal containing to-be-processed signal feature information generated by the semiconductor laser under excitation of the original optical signal; performing photoelectric conversion on the spike pulse optical signal to obtain a time domain signal; amplitude features, quantity features and time position features of spike pulses in the time domain signals are extracted, and amplitude feature virtual nodes, quantity feature virtual nodes and time position feature virtual nodes are constructed; splicing the amplitude feature virtual node, the quantity feature virtual node and the time position feature virtual node to obtain a reserve pool state vector of an extended dimension; and obtaining an output signal of the multi-dimensional pulse reservoir calculation system based on the reservoir state vector of the extended dimension.

Quantum reservoir computation is provided. A first feature vector is determined from input data. A plurality of qubits is configured in an initial configuration according to the first feature vector, wherein a detuning, Rabi frequency, phase, and / or position of each of the plurality of qubits is determined by a respective one of the values of the first feature vector. The plurality of qubits is evolved for a first time. The plurality of qubits is measured to obtain first measurements after the first time. The plurality of qubits is returned to the initial configuration. The plurality of qubits is evolved for a second time. The plurality of qubits is measured to obtain second measurements after the second time. A second feature vector is determined from the first and second measurements. The second feature vector is provided to a decoder and a characteristic of the input data is obtained therefrom.

The invention belongs to the technical field of Internet of Things. The invention provides a large-scale infrastructure network link dynamic prediction method based on nonlinear reserve pool calculation. According to the embodiment of the invention, by combining nonlinear reserve pool calculation and regularization regression technologies, the dynamic, efficient and accurate prediction of a complex network link is realized, and the method is particularly suitable for state inference and stability analysis of a key link in an infrastructure network. A reserve pool calculation framework is adopted, and the core advantage of the method is that only the linear weight of an output layer needs to be trained, and the internal connection of a huge reserve pool is kept random and fixed. A large amount of iterative computation and GPU resources required by back propagation are avoided, and the training complexity and the time cost are remarkably reduced. A Tikhonov regularization item is introduced, so that the overfitting problem under limited training data or noise interference is effectively prevented, and the generalization ability and prediction stability of the model are improved.

A method for manufacturing a reservoir computing apparatus, related to artificial intelligence. The method comprises: step a), providing a bottom electrode layer, a dielectric layer, a resistive switching layer, and a top electrode layer based on the above-listed sequence on a substrate to obtain a to-be-annealed reservoir computing apparatus; and step b), annealing the to-be-annealed reservoir computing apparatus to obtain the reservoir computing apparatus, where a temperature of the annealing ranges from 300° C. to 700° C., and duration of the annealing duration ranges from 30s to 100s. The manufactured reservoir computing apparatus is subject to rapid annealing, which redistributes defects, forms a more stable film, and introduces a ferroelectric O-phase into the film. The rapid annealing reduces power consumption and improves computing accuracy effectively.

The invention belongs to the related technical field of memory architecture design, and discloses a storage pool computing system based on reconfigurable memristors and a storage pool computing control method.The system comprises memristor-transistor units arranged in a crossed array, each memristor-transistor unit comprises a reconfigurable memristor and a transistor, and the reconfigurable memristor and the transistor are arranged in a staggered array mode. The transistor is connected with the reconfigurable memristor, the reconfigurable memristor can realize reversible conversion between volatility and non-volatility under the control of an input voltage of the transistor, the reconfigurable memristor is controlled to present volatility to realize a feature extraction process of a time sequence signal, and the reconfigurable memristor is controlled to present non-volatility to realize an identification and classification process. According to the invention, the reconfigurable memristor is utilized, and the working modes are dynamically switched through regulation and control of the transistor, so that the synergistic effect of short-term memory and long-term weight storage can be realized, and real-time and efficient reserve pool calculation can be realized; the complexity in the traditional hardware design is reduced, and the system architecture is simplified.

Disclosed in the present application are a reservoir computing network optimization method and a related apparatus. The method comprises: sampling an input signal to obtain a sampling signal; performing quantization processing on the sampling signal by means of at least two kinds of quantization modes, so as to obtain at least two kinds of digital signals, values of elements in different digital signals being different; inputting voltage pulses corresponding to the elements in the different digital signals into reservoirs constructed by different quantities of virtual nodes, so as to extract signal features of the input signal in different quantization modes by the different reservoirs. By quantizing signals in different modes and inputting same into reservoirs constructed by different quantities of virtual nodes, the richness of internal states of the reservoirs can be improved, thereby further improving the signal identification accuracy of a reservoir system.

The present invention discloses a hardware implementation method for a reservoir computing network. This method realizes the mask layer, reservoir layer, and readout layer of the reservoir computing network only through ferroelectric transistors. The mask layer, reservoir layer, and readout layer are all composed of multiple ferroelectric transistors. The gates of the ferroelectric transistors in the mask layer are connected to triangular pulses, the sources of the ferroelectric transistors in the mask layer are connected to the input voltage, the mask layer is fully connected to the reservoir layer, the gates of the ferroelectric transistors in the reservoir layer and the readout layer are both connected to the preset voltage, the reservoir layer is connected to the readout layer one by one, and the drains of all the ferroelectric transistors in the readout layer are connected in series. The current output by the connected drains is the output of the reservoir computing network. The mask layer utilizes the non-volatile storage characteristics and randomness of the ferroelectric transistors to realize the function of the mask layer. Based on the regulation of the channel current by the gates of the ferroelectric transistors, the collection of node states and the adjustment of weights are realized.

The invention discloses a next generation reserve pool calculation implementation method based on a semiconductor laser, and the method specifically comprises the steps: firstly, multiplying an input signal with a zoom factor and mask information to obtain a signal; then, the signal is subjected to intensity modulation through a Mach-Zehnder modulator, and on the premise that the length of input data is kept unchanged, time offsets of different digits are applied to the signal; directly carrying out linear superposition on each offset signal, and carrying out nonlinear operation on an obtained composite signal through a photoelectric detector to obtain a final state matrix; and finally, training the state matrix by adopting a ridge regression method so as to obtain a corresponding weight parameter. According to the method, the training efficiency is remarkably improved, the dynamic stability of the system is enhanced, the overall calculation speed is increased, and the application requirements of high speed, low power consumption, parallel calculation and the like are met.

The invention discloses a time delay reserve pool calculation method based on dispersion orthogonal polarized light feedback VCSEL. An adopted time delay RC system comprises a tunable laser, a Mach-Zehnder modulator, a VCSEL, an optical isolator, an optical coupler, a polarization controller, a variable optical attenuator, a chirp fiber Bragg grating and a photoelectric detector. A signal of the input layer is modulated by a Mach-Zehnder modulator and then is injected into an X polarization component of the VCSEL of the reserve pool layer; emitted light of the VCSEL is divided into two paths after passing through the optical circulator, one path is reflected back to the VCSEL by the chirp fiber bragg grating after being subjected to cross-polarization rotation through the polarization controller, and a reserve pool layer of the time delay RC system is constructed; one path is transmitted to the output layer and is converted into an electric signal through a photoelectric detector. According to the method, the time delay characteristic can be further inhibited, so that generation of harmful resonance is further inhibited, and the method has stronger nonlinear characteristic and better calculation performance.

This associative memory device comprises: a classification vector input data setting unit (1) that sets the same classification vector in a fixed time length as a first input vector; a time-series signal output data setting unit (2) that sets a vector composed of values at each time of a time-series signal in the time length as a first output vector; a first reservoir learning / calculation unit (3) that performs learning by reservoir computing in which a specific classification vector among a plurality of different classification vectors is associated with a specific time-series signal among a plurality of different time-series signals such that the classification vector becomes an input and the time-series signal becomes an output, and calculates a first recursive neural circuit that associates the first input vector with the first output vector; and a first reservoir circuit storage unit (4) that stores the first recursive neural circuit.

In reservoir computing, this enables calculations based on more past input information. [Solution] An information processing device that trains a model to estimate the correct output based on input has a model that includes three layers: an input layer, an intermediate layer that outputs an output value corresponding to the input value input from the input layer based on its internal state, and an output layer that outputs a value based on the output value output from the intermediate layer. In the intermediate layer, a reservoir layer, which is a collection of multiple nonlinear nodes that each perform nonlinear operations, and a linear node layer, which is a collection of multiple linear nodes that each perform linear operations, are connected in parallel to the input layer.

The present application discloses a hardware implementation method, device, and electronic device for reservoir in-memory computing, which relates to the fields of machine learning and artificial intelligence. The hardware implementation method for reservoir in-memory computing includes: obtaining a voltage signal transmitted by an input module to a memristor array; mapping a linear feature vector at a target moment to the memristor array based on the voltage signal; controlling the memristor array to determine a non-linear feature vector based on the linear feature vector at the target moment; controlling the memristor array to determine a total feature vector at the target moment based on the linear feature vector and the non-linear feature vector at the target moment; controlling the memristor array to predict the input data coordinates at the next moment based on the total feature vector at the target moment, so that an output module can convert the input data coordinates into an output value of a digital signal for output, effectively reducing the data transfer during the calculation process, improving the calculation rate of reservoir computing, and improving the reliability and stability of reservoir computing.

The application belongs to the technical field of reservoir computing, and relates to a multi-dimensional pulse reservoir computing method and system based on a semiconductor laser. The multi-dimensional pulse reservoir computing method comprises the following steps: encoding a to-be-processed signal to generate an original optical signal; injecting the original optical signal into a semiconductor laser, and collecting a spike pulse optical signal containing characteristic information of the to-be-processed signal generated by the semiconductor laser under the excitation of the original optical signal; photoelectrically converting the spike pulse optical signal to obtain a time-domain signal; extracting amplitude characteristics, quantity characteristics and time position characteristics of the spike pulse in the time-domain signal, and constructing amplitude characteristic virtual nodes, quantity characteristic virtual nodes and time position characteristic virtual nodes; splicing the amplitude characteristic virtual nodes, the quantity characteristic virtual nodes and the time position characteristic virtual nodes to obtain an extended-dimension reservoir state vector; and obtaining an output signal of the multi-dimensional pulse reservoir computing system based on the extended-dimension reservoir state vector.

The present application discloses a homogeneous optoelectronic reservoir computing system based on a nitrogen-doped Ge-Sb-Te material. The system of the present application comprises an optoelectronic reservoir layer and a readout layer which are connected to each other; the optoelectronic reservoir layer comprises a plurality of optical synaptic devices based on the nitrogen-doped Ge-Sb-Te material; the optical synaptic devices realize perception of and nonlinear response to an image optical signal on the basis of the photoconductive effect on a single optical pulse and the paired-pulse facilitation effect under dual optical pulses; the readout layer comprises a plurality of electrical synaptic devices based on the nitrogen-doped Ge-Sb-Te material; the electrical synaptic devices realize linear response to and image recognition of an output signal of the optoelectronic reservoir layer on the basis of the linearity and the symmetric long-term potentiation and long-term depression functions. In the system of the present application, the reservoir layer and the readout layer use the devices based on the same material, thereby realizing the homogeneous optoelectronic reservoir computing system and higher system integration and process compatibility.

The invention relates to a signal recovery method, system and device based on a space-time interconnection photoelectric reservoir computing architecture and a medium, and the method comprises the steps: constructing a photoelectric reservoir network based on small time delay feedback: introducing one path of a modulation optical signal into a time delay feedback loop, and forming a photoelectric oscillator; the total delay of the time delay feedback loop is divided into a plurality of time slots at equal intervals, each time slot serves as a virtual node, and a time division multiplexing reservoir state matrix is constructed in a time domain; sampling an output signal of the photoelectric oscillator, and performing serial-parallel conversion to output K paths of spatial parallel optical signals; inputting the spatial parallel optical signals into an integrated space photon processing platform, and performing product operation of a reservoir state matrix and an output weight matrix; and receiving the optical signal output by the integrated space photon processing platform in parallel and converting the optical signal into an electric signal to obtain a prediction output result of the target signal. Compared with the prior art, the system has the advantages of high speed, low power consumption and high expansibility.

The present invention discloses an online modeling and prediction method for the state information of an unmanned ship based on simulation technology, which relates to the technical field of unmanned ship simulation testing. A database system for the unmanned ship is constructed, including simulation data obtained by simulating an unmanned ship model in a real flow field of star-ccm+ and a Noetic ocean environment, as well as data collected during real sea trials. A nonlinear reservoir computing algorithm based on a sliding window algorithm is designed to perform online modeling and prediction of the state information. By combining the dynamic characteristics of computational fluid dynamics, physical simulation environment, and real ocean environment, the present invention can enable the unmanned ship to realize online modeling and prediction of state information using small sample data in the real ocean environment, provide effective prior information for collision avoidance, planning, and control of the unmanned ship, and safely and effectively complete the navigation task.

The invention provides a reservoir calculation method based on rare earth doped nanocrystals. The reservoir calculation method comprises the following steps: acquiring an original digital image and converting the original digital image into a binary vector set; modulating the binary vector set into a pulse sequence set through time multiplexing coding; inputting the pulse sequence into a reservoir formed by three layers of core-shell-shell structure rare earth doped nanocrystals, and generating a corresponding luminescence response curve set through nonlinear dynamic interaction of the three layers of core-shell-shell structure rare earth doped nanocrystals; extracting features based on the curve set to form a reservoir state vector set; and finally, training and classifying the state vector set through a linear layer to obtain an image classification result. According to the reservoir calculation method based on the rare earth doped nanocrystals, through the synergistic effect of the physical characteristics of the rare earth doped nanocrystals and time multiplexing coding, a single reservoir replaces a traditional complex electronic network, coding, calculation and feature extraction are completed in an optical domain, the algorithm complexity and the training cost are remarkably reduced, and the calculation efficiency is improved. And efficient image classification is realized.

The application relates to a discrete dislocation dynamics acceleration simulation method based on reservoir computing, which effectively reduces the high computing cost caused by the global time step limit and the long-range interaction of dislocation elastic force in traditional dislocation dynamics simulation by replacing the node motion speed calculation in discrete dislocation dynamics with a trained reservoir computing model. The virtual node mechanism is innovatively proposed to solve the dimension mismatch problem caused by the dynamic change of the number of nodes in the dislocation topology evolution process, and adaptive discretization prediction is realized. The prediction error accumulation is prevented by regularly inserting the real traditional dislocation dynamics simulation calculation result in the prediction sequence, and the stability of long-term simulation is ensured. The acceleration scheme significantly reduces the computing time complexity from exponential to linear while maintaining the physical accuracy, providing a feasible technical path for efficient simulation of large-scale three-dimensional complex dislocation network evolution.

The invention relates to a database operation method and device for function calculation, computer equipment, a computer readable storage medium and a computer program product. The method applied to any function computing node in a function computing cluster comprises the steps that a computing layer in a current function computing node receives a function computing request, wherein the function computing request comprises an object identifier and a function identifier; if the data of the object identifier is stored in the storage layer of the other function computing node, the computing layer forwards the function computing request to the other function computing node; if the data is stored in the storage layer of the current function calculation node, the calculation layer starts the database transaction, the function instance is executed based on the data of the object identifier, and when the function is executed, reading operation is performed on the distributed storage database provided by the storage layer through the database transaction, and data required by writing operation on the distributed storage database is cached; and after the function execution is finished, the database transaction is closed, the cached data is submitted to a storage layer for write operation, and the communication overhead is reduced.

The application belongs to the technical field of neuromorphic computing devices, and particularly relates to a temperature programmable neuromorphic device based on lead sulfide colloidal quantum dots and application. The device comprises an insulating substrate, a planar gap electrode, a lead sulfide colloidal quantum dot photosensitive film and a temperature control module. By setting the working temperature to control the release kinetics of the trap state charge in the photosensitive film, the device has different photocurrent decay characteristics at different temperatures, realizing reversible switching of the fast detection and long-term storage modes. The device can be used for reservoir computing, and realizes time sequence information processing through the historical dependence response to time sequence light pulses. The application solves the problem of fixed and unadjustable response time scale of the existing device, and has the advantages of simple structure, good process compatibility and wide response dynamic range.

Various examples related to reservoir computing are provided. In one example, a physical reservoir computer includes a processing circuit, the processing circuit having an input layer, a reservoir implemented with a forced limit cycle oscillator and implemented without delay or feedback, and a readout layer. The forced limit cycle oscillator can include a Hopf oscillator or a Lorenz oscillator. The processing circuit can include an analog processing circuit, an optoelectronic circuit, or other suitable processing circuit.

One technical solution of the present invention is to provide a deep reservoir optical computing method based on semiconductor lasers, and another technical solution of the present invention is to provide a deep reservoir optical computing system. In the present invention, each layer of the reservoir is mainly composed of a semiconductor laser and an optical delay line, and a large number of virtual neurons are generated thereby. The output of the laser in the upper layer of the reservoir is unidirectionally injected into the laser of the lower layer of the reservoir through the optical injection locking technology, thereby realizing the deep architecture of the reservoir computing. The depth of the reservoir optical computing system proposed by the present invention is not limited, so it has good scalability. The connection between the reservoir layers is an all-optical connection, so it has the advantages of simple device, low cost and low energy consumption.

The present application provides a method, apparatus, device, and signal equalizer based on reservoir computing. The method includes: when a predetermined trigger condition is satisfied, generating neural network parameter information corresponding to a second communication device according to channel condition information reported by the second communication device, where the neural network parameter information includes a jump size corresponding to a reservoir; and sending the neural network parameter information to the second communication device. The present application proposes the concept of DRJ. The second communication device can configure DRJ based on the jump size allocated by the first communication device for it, so as to implement a signal equalizer based on DRJ, which can support adaptive equalization of a high-speed communication system in an easy-to-train and low-power consumption manner. Due to the structural characteristics brought by its jump-carrying deterministic structure and relatively simple linear regression training characteristics, it can reduce the complexity and power consumption of the second communication device and enable it to support a simpler 、 Lower power consumption and faster hardware implementation.

The invention relates to the technical field of memristor array design, and discloses a 3D stacked memristor array design method for high-energy-efficiency AI calculation, and the method comprises the following steps: S1, constructing a layered heterogeneous 3D stacked architecture which comprises the physical isolation and dynamic reconstruction of a storage layer, a calculation layer and a control layer; s2, establishing a coupling relation model of a temperature field and memristor resistance state drift based on a dynamic thermal-electric cooperative regulation and control algorithm, and optimizing voltage pulse parameters in real time; s3, dynamically distributing mixed precision pulses according to the sparse characteristic of the AI model through sparse perception adaptive gating; and S4, deploying a hardware acceleration unit in a control layer, and realizing thermal-electric parameter calculation and real-time coding transmission of a regulation and control instruction. By adopting the technical scheme of layered heterogeneous 3D stacking architecture and hybrid bonding vertical interconnection, the effect of data near processing and transmission loss collaborative optimization is achieved, and the problems of extra energy consumption and delay caused by cross-layer data handling are solved.

A reservoir computing method (700) and related system (100) are disclosed. The reservoir computing system comprises at least one nonlinear compute node (103) with compute node MOSFET (113), a driver circuit(110) coupled to the at least one nonlinear compute node (103), and a readout unit (104) for acquiring an output sequence related to the at least one nonlinear compute node (103). The MOSFET has charge trapping sites located in a gate dielectric and displays bias temperature instability when subjected to voltage stress and threshold voltage recovery after removal of the voltage stress. The driver unit is configured to apply input pulses and readout pulses to a gate of the MOSFET. The readout unit (104) is configured to detect a current through a channel region of the MOSFET while the driver circuit is applying the readout pulses. The readout pulses have a smaller voltage amplitude compared to the input pulses.

The invention relates to a discrete dislocation dynamics acceleration simulation method based on reservoir calculation, and the method comprises the steps: replacing the node movement speed calculation in discrete dislocation dynamics with a trained reservoir calculation model; the high calculation cost caused by global time step length limitation and dislocation elastic force long-range interaction in traditional dislocation dynamics simulation is effectively reduced. A virtual node mechanism is innovatively provided, the problem of dimension mismatching caused by dynamic change of the number of nodes in the dislocation topology evolution process is solved, and self-adaptive discretization prediction is achieved. And real traditional dislocation dynamics simulation calculation results are periodically inserted into the prediction sequence, so that prediction error accumulation is prevented, and the stability of long-term simulation is ensured. According to the acceleration scheme, while the physical accuracy is kept, the calculation is obviously reduced from the index time complexity to the linear time complexity, and a feasible technical path is provided for efficient simulation of large-scale three-dimensional complex dislocation network evolution.

The invention discloses a method for realizing reserve pool calculation based on organoid, which comprises the following steps: acquiring an audio signal, and converting the acquired audio signal into a photoelectric stimulation pulse signal; acquiring a fluorescence signal responding to the photoelectric stimulation pulse signal through the organ-like device; according to the obtained fluorescence signals, high-dimensional feature data are obtained through calculation by using a storage pool; and decoding the high-dimensional feature data to complete identification processing of the audio signal. According to the method, a biological organoid is adopted as a core of a reserve pool, energy consumption is extremely low when the organoid executes perception and cognition tasks depending on the natural storage and calculation fusion characteristic of a biological neural network, organic fusion of data storage and processing is achieved through neural synapses, a large amount of energy consumption of CPU and memory data transmission in a traditional von Neumann architecture is avoided, and the reliability of the von Neumann architecture is improved. And the development requirement of low-power AI hardware is better met, and a thought is provided for developing an efficient and low-cost neuromorphic chip.

A reservoir computing method (700) and related system (100) are disclosed. The reservoir computing system comprises at least one nonlinear compute node (103) with compute node MOSFET (113), a driver circuit (110) coupled to the at least one nonlinear compute node (103), and a readout unit (104) for acquiring an output sequence related to the at least one nonlinear compute node (103). The MOSFET has charge trapping sites located in a gate dielectric and displays bias temperature instability when subjected to voltage stress and threshold voltage recovery after removal of the voltage stress. The driver unit is configured to apply input pulses and readout pulses to a gate of the MOSFET. The readout unit (104) is configured to detect a current through a channel region of the MOSFET while the driver circuit is applying the readout pulses. The readout pulses have a smaller voltage amplitude compared to the input pulses.

A system for device authentication in Industrial Internet of Things (IIoT) networks, consisting of: a device behavior monitoring module configured to capture device communication patterns, including time intervals between packets, protocol usage, packet sizes, and communication frequency; a multi-persistence homology processor configured to extract topological features from the captured device communication patterns across multiple timescales and generate persistent features representing device behavior; a cognitive behavioral profiling engine configured to generate cognitive fingerprints based on the device's communication patterns and compute cognitive signatures as multi-persistence embeddings, with the developed cognitive behavioral signatures being stored in a database; an embedding generator configured to compress the topological features and cognitive signatures into low-dimensional behavioral signatures using at least one autoencoder and reservoir computing technique; an authentication module configured to compare current behavioral signatures with behavioral signatures stored in the database and to authenticate devices based on the signature match; and a reinforcement learning module configured to adaptively update authentication thresholds based on real-time device behavior and optimize decision parameters through feedback learning.