Complex network generation method and apparatus, computer device, and storage medium

Abstract

The application relates to a complex network generation method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: determining a node arrangement structure, node distances between spatial nodes and node connection constraints; the spatial nodes comprise nodes at different positions; determining a distance lower limit value and a distance upper limit value based on a decay time length of a memristor and a first interval time length; the first interval time length supports dynamic adjustment; sampling a boundary distance of the spatial nodes based on a first target probability density function to obtain a target boundary distance; the target boundary distance is located between the distance lower limit value and the distance upper limit value; selecting other spatial nodes with a node distance smaller than the target boundary distance from the spatial nodes as adjacent nodes of the spatial nodes; and connecting the spatial nodes and the adjacent nodes of the spatial nodes according to the node connection constraints in the node arrangement structure to obtain a target complex network. The method can improve the flexibility of the network structure.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a complex network generation method and device, computer equipment and a storage medium. BACKGROUND

[0002] Connectionist artificial intelligence inspired by brain neural connections, such as deep learning, has achieved unprecedented success in the past decade. The goal of artificial neural network design and optimization gradually transitions from training connection weights to designing network structures. However, this trend is lagging behind the large hardware overhead of traditional CMOS neuron circuits, which makes the scale of ANN (Artificial Neural Network) accelerators or neuromorphic chips cannot approach the scale of the human brain, although it is partially solved by using scalable memristors with more general neuromorphic functions, but it can only provide limited input and output and connection number. Obviously, the current artificial neural network has the problem of poor network structure flexibility. SUMMARY

[0003] Therefore, it is necessary to provide a complex network generation method, device, computer equipment, computer readable storage medium and computer program product capable of improving the flexibility of the network structure to solve the above technical problems.

[0004] In a first aspect, the present application provides a complex network generation method. The method comprises:

[0005] determining a node arrangement structure, a node distance between space nodes, and a node connection constraint; the space nodes include nodes at different positions;

[0006] determining a distance lower limit value and a distance upper limit value based on a decay duration of a memristor and a first interval duration; the first interval duration supports dynamic adjustment;

[0007] sampling a boundary distance of the space nodes based on a first target probability density function to obtain a target boundary distance; the target boundary distance is located between the distance lower limit value and the distance upper limit value;

[0008] selecting other space nodes with a node distance from the space nodes less than the target boundary distance as adjacent nodes of the space nodes;

[0009] connecting the space nodes and the adjacent nodes of the space nodes according to the node connection constraint under the node arrangement structure to obtain a target complex network.

[0010] In one embodiment, the node arrangement structure includes a circular ring arrangement structure; and the determining the node arrangement structure, the node distance between the space nodes, and the node connection constraint comprises:

[0011] obtaining the circular arrangement structure;

[0012] arranging the space nodes according to the circular arrangement structure;

[0013] obtaining a clockwise connection rule and an open loop connection rule;

[0014] taking the clockwise connection rule and the open loop connection rule as node connection constraints.

[0015] In one embodiment, the method further comprises:

[0016] obtaining an initial probability density function;

[0017] determining an expected value and a variance based on the distance lower limit value and the distance upper limit value;

[0018] substituting a coefficient parameter, the variance and the expected value into the initial probability density function to obtain a probability density function;

[0019] generating a key constant based on the probability density function, the distance lower limit value and the distance upper limit value;

[0020] combining the probability density function and the key constant to obtain a first target probability density function.

[0021] In one embodiment, the selecting other space nodes with a node distance less than the target boundary distance from the space node as adjacent nodes of the space node comprises:

[0022] determining the node distance between the space node and other space nodes in a clockwise direction;

[0023] when the node distance is less than the target boundary distance, taking the other space node corresponding to the node distance as an adjacent node of the space node;

[0024] The method further comprises:

[0025] when the node distance is not less than the target boundary distance, taking the other space node corresponding to the node distance as a remote node of the space node; the remote node is not connected to the node.

[0026] In one embodiment, the connecting the space node and the adjacent node of the space node according to the node connection constraints under the node arrangement structure to obtain a target complex network comprises:

[0027] Under the node arrangement structure, the space nodes and the adjacent nodes of the space nodes are connected according to the node connection constraint, so as to obtain an initial complex network;

[0028] obtaining a network quantization index; the network quantization index includes a clustering coefficient, a characteristic path length and a small-world coefficient;

[0029] adjusting the initial complex network according to the network quantization index, so as to obtain a target complex network.

[0030] In one of the embodiments, the method further comprises:

[0031] If the second interval duration of the memristor is not greater than the lower limit of the decay duration, and the product of the first node number and the second interval duration is not less than the upper limit of the decay duration, then the decay boundary time of the time node of the first node number is sampled based on a second target probability density function, so as to obtain a target boundary time; the time node includes nodes at different time points; the target boundary time is located between the lower limit of the decay duration and the upper limit of the decay duration; and the second interval duration supports dynamic adjustment;

[0032] Selecting other time nodes with a second interval duration less than the target boundary time as adjacent nodes of the time node;

[0033] connecting the time node and the adjacent nodes of the time node, so as to obtain a new reserve pool layer; the new reserve pool layer generates a reserve pool calculation model.

[0034] In one of the embodiments, the reserve pool calculation model is a same-type calculation model or a hybrid calculation model; the method further comprises:

[0035] obtaining a short-term memory task or a parity check task, and executing the short-term memory task or the parity check task through the same-type calculation model; the same-type calculation model is generated based on at least one same-type memristor;

[0036] obtaining a speech digit recognition task, and executing the speech digit recognition task through the hybrid calculation model; the hybrid calculation model is generated based on at least two types of memristors.

[0037] In a second aspect, the application further provides a complex network generation device. The device comprises:

[0038] a first determination module for determining a node arrangement structure, node distances between space nodes and node connection constraints; the space nodes include nodes at different positions;

[0039] The second determining module is configured to determine a distance lower limit value and a distance upper limit value based on the decay time length of the memristor and a first interval time length; and the first interval time length supports dynamic adjustment.

[0040] The sampling module is configured to sample a boundary distance of the spatial node based on the first target probability density function, to obtain a target boundary distance; and the target boundary distance is located between the distance lower limit value and the distance upper limit value.

[0041] The selecting module is configured to select other spatial nodes with a node distance to the spatial node less than the target boundary distance as neighboring nodes of the spatial node.

[0042] The connecting module is configured to connect the spatial node and the neighboring nodes of the spatial node according to the node connection constraint under the node arrangement structure, to obtain a target complex network.

[0043] In one of the embodiments, the node arrangement structure includes a circular ring arrangement structure; the first determining module is further configured to obtain the circular ring arrangement structure; after arranging the spatial nodes according to the circular ring arrangement structure, the distances between the spatial nodes are determined in a clockwise direction; a clockwise connection rule and an open loop connection rule are obtained; and the clockwise connection rule and the open loop connection rule are taken as the node connection constraint.

[0044] In one of the embodiments, the device further includes:

[0045] The processing module is configured to obtain an initial probability density function; determine an expected value and a variance based on the distance lower limit value and the distance upper limit value; substitute a coefficient parameter, the variance and the expected value into the initial probability density function, to obtain a probability density function; generate a key constant based on the probability density function, the distance lower limit value and the distance upper limit value; and combine the probability density function and the key constant to obtain a first target probability density function.

[0046] In one of the embodiments, the selecting module is further configured to determine a node distance between the spatial node and other spatial nodes in a clockwise direction; when the node distance is less than the target boundary distance, the other spatial node corresponding to the node distance is taken as a neighboring node of the spatial node; when the node distance is not less than the target boundary distance, the other spatial node corresponding to the node distance is taken as a distal node of the spatial node; and the distal node is not connected to the node.

[0047] In one of the embodiments, the connecting module is further configured to connect the spatial nodes and the adjacent nodes of the spatial nodes according to the node connection constraint under the node arrangement structure, to obtain an initial complex network; obtain network quantization indexes; the network quantization indexes include a clustering coefficient, a characteristic path length, and a small-world coefficient; and adjust the initial complex network according to the network quantization indexes, to obtain a target complex network.

[0048] In one of the embodiments, the sampling module is further configured to sample a target boundary time of the first node number of time nodes based on a second target probability density function, if the second interval time length of the memristor is not greater than a lower limit value of the decay time length, and the product of the first node number and the second interval time length is not less than an upper limit value of the decay time length; the time nodes include nodes at different time instants; the target boundary time is located between the lower limit value of the decay time length and the upper limit value of the decay time length; and the second interval time length supports dynamic adjustment; the selecting module is further configured to select other time nodes with a second interval time length smaller than the target boundary time as adjacent nodes of the time nodes; and the connecting module is further configured to connect the time nodes and the adjacent nodes of the time nodes, to obtain a new reserve pool layer; and the new reserve pool layer generates a reserve pool calculation model.

[0049] In one of the embodiments, the reserve pool calculation model is a same-type calculation model or a hybrid calculation model; and the device further includes:

[0050] The executing module is further configured to obtain a short-term memory task or a parity check task, and execute the short-term memory task or the parity check task through the same-type calculation model; the same-type calculation model is generated based on at least one same-type memristor; obtain a speech digit recognition task, and execute the speech digit recognition task through the hybrid calculation model; and the hybrid calculation model is generated based on at least two types of memristors.

[0051] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0052] In a fourth aspect, the present application further provides a computer readable storage medium. The computer readable storage medium stores a computer program, and the computer program implements the steps of the above method when executed by a processor.

[0053] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and the computer program implements the steps of the above method when executed by a processor.

[0054] The complex network generation method, device, computer device, storage medium and computer program product, by determining the node arrangement structure, the node distance between the space nodes and the node connection constraint; the space nodes include nodes at different positions; the distance lower limit value and the distance upper limit value are determined based on the decay duration of the memristor and the first interval duration; the first interval duration supports dynamic adjustment; by dynamically adjusting the first interval duration, the adjustment of the network structure is realized, the reconfigurability of the generated complex network at the level of the memristor is realized, and the flexibility of the network structure is effectively improved. Moreover, by sampling the boundary distance of the space node based on the first target probability density function, a target boundary distance is obtained; the target boundary distance is between the distance lower limit value and the distance upper limit value; other space nodes with a node distance less than the target boundary distance from the space node are selected as adjacent nodes of the space node; under the node arrangement structure, the space nodes and the adjacent nodes of the space nodes are connected according to the node connection constraint, to obtain a target complex network, the target boundary distance is obtained by sampling the first target probability density function, and the adjacent nodes determined according to the obtained target boundary distance are connected, realizing the flexibility of the connection between the nodes, thereby further improving the flexibility of the network structure while ensuring the network computing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 An application environment diagram of the complex network generation method in an embodiment;

[0056] Figure 2 A flowchart of the complex network generation method in an embodiment;

[0057] Figure 3 A flowchart of the step of generating a reserve pool computing model in an embodiment;

[0058] Figure 4 An electrical characteristic diagram of a memristor in an embodiment;

[0059] Figure 5 A statistical analysis diagram of the decay duration of a memristor in an embodiment;

[0060] Figure 6 A schematic diagram of a complex network in an embodiment;

[0061] Figure 7 A schematic diagram of a network quantization index in an embodiment;

[0062] Figure 8 A performance comparison diagram of a target complex network and a typical complex network in an embodiment;

[0063] Figure 9 A schematic diagram of a new reserve pool layer in an embodiment;

[0064] Figure 10 a related schematic diagram of a reserve pool calculation model in one embodiment;

[0065] Figure 11 a structural block diagram of a complex network generation device in one embodiment;

[0066] Figure 12 a structural block diagram of a complex network generation device in one embodiment;

[0067] Figure 13 an internal structure diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0068] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0069] The complex network generation method provided by the embodiments of the present application can be applied to an application environment as shown in the accompanying drawings. Figure 1 As shown in the accompanying drawings, the terminal 102 communicates with the server 104 through a network. The data storage system can store data required to be processed by the server 104. The data storage system can be integrated on the server 104, or placed on a cloud or other network server. The present application can be executed by the terminal 102 or the server 104, and the embodiments are described by taking the terminal 102 as an example.

[0070] The terminal 102 determines a node arrangement structure, node distances between space nodes, and node connection constraints; the space nodes include nodes at different positions; the terminal 102 determines a distance lower limit value and a distance upper limit value based on a decay time length of a memristor and a first interval time length; the first interval time length supports dynamic adjustment; the terminal 102 samples a boundary distance of the space nodes based on a first target probability density function, to obtain a target boundary distance; the target boundary distance is located between the distance lower limit value and the distance upper limit value; and the terminal 102 selects other space nodes with a node distance to the space nodes less than the target boundary distance as adjacent nodes of the space nodes.

[0071] The terminal 102 connects the space nodes and the adjacent nodes of the space nodes according to the node connection constraints under the node arrangement structure, to obtain a target complex network.

[0072] The terminal 102 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The server 104 can be implemented by a stand-alone server or a server cluster composed of multiple servers.

[0073] In one embodiment, as shown in FIG. 1, a complex network generation method is provided. The method is applied to the terminal 102 in FIG. 1 as an example for illustration, including the following steps: Figure 2 Figure 1 In one embodiment, as shown in FIG. 1, a complex network generation method is provided. The method is applied to the terminal 102 in FIG. 1 as an example for illustration, including the following steps:

[0074] S202, determining a node arrangement structure, a node distance between space nodes, and a node connection constraint; the space nodes include nodes at different positions.

[0075] The node arrangement structure can refer to the arrangement structure of the nodes, and the node arrangement structure includes, but is not limited to, a circular ring arrangement structure, a rectangular arrangement structure, a trapezoidal arrangement structure, and a triangular arrangement structure. The space nodes can refer to nodes generated based on different positions in space, and the space nodes include nodes at different positions. The node distance can refer to the distance between the nodes. The node connection constraint can be used to determine the connection mode between the nodes, and the node connection constraint includes a clockwise connection rule and an open loop connection rule. The clockwise connection rule can refer to connecting the nodes in a clockwise direction. The open loop connection rule can refer to that if there is an open loop gap between two nodes arranged in a clockwise direction, the latter node is not an adjacent node of the former node.

[0076] Specifically, the node arrangement structure includes a circular ring arrangement structure; the circular ring arrangement structure is obtained; after the space nodes are arranged according to the circular ring arrangement structure, the node distance between the space nodes is determined in a clockwise direction; the clockwise connection rule and the open loop connection rule are obtained; and the clockwise connection rule and the open loop connection rule are taken as the node connection constraint.

[0077] S204, determining a distance lower limit value and a distance upper limit value based on a decay time length of the memristor and a first interval time length; the first interval time length supports dynamic adjustment.

[0078] The memristor includes a volatile memristor and a non-volatile memristor. The volatile memristor can also be referred to as a dynamic memristor. The memristor in this application can be a volatile memristor. The volatile memristor can switch resistance in response to electrical stimulation, and in the absence of external stimulation, they will gradually return to the original state over time, exhibiting time dynamic behavior. The dynamic memristor has a vertical stack of Ta / Ta2O x ​A cross-point structure of / HfO2 / Pd (50 nm / 15 nm / 15 nm / 50 nm). The decay time (τ) can refer to a time length for the memristor to recover to an original state from stopping electrical stimulation. The interval time can refer to an interval time (θ) for electrical stimulation of the memristor. The first interval time and the second interval time are different interval times. The distance lower limit value can refer to a ratio between a lower limit value of the decay time of the memristor and the first interval time, and the distance lower limit value (D min ) is calculated by a formula: D min = τ min / θ. The distance upper limit value can refer to a ratio between an upper limit value of the decay time of the memristor and the first interval time, and the distance upper limit value (D min ) is calculated by a formula: D max = τ max / θ.

[0079] In S206, a boundary distance of the spatial node is sampled based on the first target probability density function to obtain a target boundary distance; the target boundary distance is located between the distance lower limit value and the distance upper limit value.

[0080] The first target probability density function can refer to a function for sampling the boundary distance. The boundary distance can be used to determine the target boundary distance. The target boundary distance can be used to determine the proximal node and the distal node of the spatial node.

[0081] In one embodiment, an initial probability density function is obtained; an expectation value and a variance are determined based on the distance lower limit value and the distance upper limit value; a coefficient parameter, the variance and the expectation value are substituted into the initial probability density function to obtain a probability density function; a key constant is generated based on the probability density function, the distance lower limit value and the distance upper limit value; and the probability density function and the key constant are combined to obtain the first target probability density function.

[0082] The initial probability density function can be used as an initial function for generating the first target probability density function. The expectation value can refer to a value of an expectation parameter in the initial probability density function. The variance can refer to a value of a variance parameter in the initial probability density function. The coefficient parameter can refer to a coefficient in the initial probability density function. The probability density function can refer to a function obtained by substituting the coefficient parameter, the variance and the expectation value into the initial probability density function. The key constant can refer to a constant obtained based on the probability density function, the distance lower limit value and the distance upper limit value.

[0083] In one embodiment, determining the expectation value and the variance based on the distance lower limit value and the distance upper limit value includes: substituting the distance upper limit value and the distance lower limit value into an expectation function to obtain the expectation value, and substituting the distance upper limit value and the distance lower limit value into a variance function to obtain the variance.

[0084] In one embodiment, the coefficient parameter, the variance and the expectation value are substituted into the initial probability density function to obtain the probability density function, including: obtaining a standard deviation according to the variance, and substituting the coefficient parameter, the variance, the standard deviation and the expectation value into the initial probability density function to obtain the probability density function.

[0085] In one embodiment, the key constant is generated based on the probability density function, the distance lower limit value and the distance upper limit value, including: after substituting the distance lower limit value and the distance upper limit value into the probability density function, performing integral processing to obtain the key constant.

[0086] In one embodiment, the combination of the probability density function and the key constant to obtain the first target probability density function includes adding the probability density function and the key constant to obtain the first target probability density function.

[0087] For example, the initial probability density function is

[0088]

[0089] Wherein, the system parameter A = 15.5, the expectation value The standard deviation The variance The key constant D min And D max are two limits of D i , and the probability becomes zero beyond the limit. Between D min and D max , the probability distribution function is a Gaussian function vertically translated by ε, which ensures the normalization of the probability of the entire sample space.

[0090] S208, selecting other space nodes with a node distance from the space node less than a target boundary distance as adjacent nodes of the space node.

[0091] Wherein, the other space nodes can refer to space nodes other than the current space node. The adjacent nodes refer to nodes that can be connected.

[0092] In one embodiment, the node distance between the space node and the other space nodes is determined in a clockwise direction; when the node distance is less than the target boundary distance, the other space node corresponding to the node distance is selected as the adjacent node of the space node, and when the node distance is not less than the target boundary distance, the other space node corresponding to the node distance is selected as the remote node of the space node; the remote node is not connected to the space node. Wherein, the remote node refers to a node that cannot be connected.

[0093] For example, in a clockwise direction, there are three nodes in sequence, the spatial node A, the spatial node B and the spatial node C, the target boundary distance of the spatial node A is 3, the target boundary distance of the spatial node B is 4, the target boundary distance of the spatial node C is 3, the node distance between A and B is 2, the node distance between B and C is 4, and the node distance between A and C is 3. It can be known that the spatial node B is the adjacent node of the spatial node A, which has a node distance less than the target boundary distance 3 of the spatial node A. The spatial node A is the adjacent node of the spatial node B, which has a node distance less than the target boundary distance 4 of the spatial node B. There is no spatial node having a node distance less than the target boundary distance 3 of the spatial node C.

[0094] S210, under the node arrangement structure, connecting the spatial nodes and the adjacent nodes of the spatial nodes according to the node connection constraint to obtain a target complex network.

[0095] The target complex network can refer to a target obtained complex network.

[0096] In one embodiment, under the circular ring arrangement structure, the adjacent nodes of the spatial nodes are sequentially screened according to the clockwise direction determined in the clockwise connection rule to obtain screened adjacent nodes, and the adjacent nodes of the spatial nodes are connected with the screened adjacent nodes according to the open loop connection rule to obtain the target complex network.

[0097] In one embodiment, under the node arrangement structure, the spatial nodes and the adjacent nodes of the spatial nodes are connected according to the node connection constraint to obtain an initial complex network; a network quantitative index is obtained; the network quantitative index includes a clustering coefficient, a characteristic path length and a small world coefficient; the initial complex network is adjusted according to the network quantitative index to obtain the target complex network.

[0098] The initial complex network can refer to an initial complex network in a training process. The network quantitative index can refer to an index for measuring the performance of a complex network. The network quantitative index includes a clustering coefficient (C*), a characteristic path length (L*), a small world coefficient (S*), a global efficiency (GE), a local efficiency (LE), a tree width (T), a maximum clique size (MCS), a radius (R) and a wiring cost (E).

[0099] In one embodiment, the initial complex network is adjusted according to a network quantification index to obtain a target complex network, including: adjusting the initial complex network according to at least one index of a clustering coefficient (C*), a characteristic path length (L*), a small-world coefficient (S*), a global efficiency (GE), a local efficiency (LE), a tree width (T), a maximum clique size (MCS), a radius (R), and a wiring cost (E) to obtain the target complex network.

[0100] In the complex network generation method, the node arrangement structure, the node distance between the space nodes, and the node connection constraint are determined; the space nodes include nodes at different positions; the distance lower limit value and the distance upper limit value are determined based on the decay duration of the memristor and the first interval duration; the first interval duration supports dynamic adjustment; the network structure is adjusted by dynamically adjusting the first interval duration, the reconfigurability of the generated complex network at the level of the memristor is achieved, and the flexibility of the network structure is effectively improved. Moreover, the target boundary distance is obtained by sampling the boundary distance of the space node based on the first target probability density function; the target boundary distance is located between the distance lower limit value and the distance upper limit value; other space nodes with a node distance less than the target boundary distance from the space node are selected as adjacent nodes of the space node; under the node arrangement structure, the space node and the adjacent nodes of the space node are connected according to the node connection constraint to obtain a target complex network, the target boundary distance is obtained by sampling the first target probability density function, and the adjacent nodes determined according to the obtained target boundary distance are connected, the flexibility of the connection between nodes is achieved, and the flexibility of the network structure is further improved while ensuring the network computing efficiency.

[0101] In one embodiment, as shown in Figure 3 the reserve pool computing model generation step includes:

[0102] S302, if the second interval duration of the memristor is not greater than the decay duration lower limit value, and the product of the first node quantity and the second interval duration is not less than the decay duration upper limit value, the decay boundary time of the time node of the first node quantity is sampled based on the second target probability density function to obtain a target boundary time; the time node includes nodes at different time instants; the target boundary time is located between the decay duration lower limit value and the decay duration upper limit value; the second interval duration supports dynamic adjustment;

[0103] wherein the second interval duration can refer to an interval duration (θ) different from the first interval duration. The decay duration lower limit value (τ min ) can refer to the minimum value of the decay duration of the memristor. The decay duration upper limit value (τ max) can refer to the maximum value of the decay duration of the memristor. The first node number can refer to the node number of the time node. The second target probability density function can refer to a function sampled on the decay boundary time, and the generation process of the second target probability density function can be analogous to the generation process of the first target probability density function, that is, the generation process of the second target probability density function can refer to the above-mentioned embodiments of the generation of the first target probability density function. The time node can refer to a node generated based on different time points in the time domain, and the time node includes nodes at different time points. The decay boundary time can be used to determine the target boundary time. The target boundary time can be used to determine the adjacent node and the remote node of the time node.

[0104] S304, selecting other time nodes with a second interval duration smaller than the target boundary time from the time node as adjacent nodes of the time node.

[0105] Among them, the other time node can refer to a time node other than the current time node.

[0106] In one embodiment, according to the time sequence, the second interval time between each time node and other time nodes is determined in turn, when the second interval time is smaller than the target boundary time, the other time node corresponding to the second interval time is taken as the adjacent node of the time node, when the second interval time is not smaller than the target boundary time, the other time node corresponding to the second interval time is taken as the remote node of the time node; the remote node is not connected with the time node.

[0107] For example, according to the time sequence, there are three nodes in turn, the time node A, the time node B and the time node C, assuming that the second interval time between adjacent time nodes is θ, that is, the second interval time between the time node A and the time node B is θ, the second interval time between the time node A and the time node C is 2θ, the second interval time between the time node B and the time node C is θ, the target boundary time of the time node A is τ1, the target boundary time of the time node B is τ2, and the target boundary time of the time node C is τ3, then the time node with the second interval time smaller than the target boundary time τ1 of the space node A is the adjacent node of the time node A; the time node with the second interval time smaller than the target boundary time τ2 of the space node B is the adjacent node of the time node B; the time node with the second interval time smaller than the target boundary time τ3 of the space node C is the adjacent node of the time node C.

[0108] S306, connecting the time node and the adjacent node of the time node to obtain a new reserve pool layer; the new reserve pool layer generates a reserve pool calculation model.

[0109] The new reservoir layer can also be referred to as a new pool layer, and can refer to a reservoir layer generated in the present application. The new reservoir layer is used to generate a reservoir computing model. The reservoir computing model (RC) model can also be referred to as a reservoir, which refers to an efficient artificial neural network suitable for processing time series signals. The RC model is one of the most advanced connection models, and is a high-dimensional nonlinear dynamic system in which the input is nonlinearly converted into a high-dimensional state space, and different inputs are more easily separated. Therefore, the RC model is usually modeled by a high-dimensional state vector, such as a neural network, in which the state of each neuron is considered as a component of the reservoir state vector. Unlike mainstream deep learning, RC relaxes the requirement for precise adjustment of synaptic connections, and precise weight distribution seems to have limited impact on its performance. In order to increase the dimension of the reservoir response, a large number of neurons are required. In order to reduce the number of neurons in the spatial domain, Appeltant et al. proposed using time division multiplexing to create a high-dimensional reservoir with a single dynamic node. Unlike creating multiple spatially separated neurons, a single dynamic node is multiplexed and reused sequentially, creating a new virtual node in each multiplexing period, which plays a similar role to a neuron. Due to the mitigation of large-scale interconnection and excessive hardware overhead problems, such reservoir layers have been widely applied in electronic and photonic hardware, including memristor devices.

[0110] In one embodiment, the reservoir computing model obtains a short-term memory task or a parity check task through a homogeneous computing model, and executes the short-term memory task or the parity check task through the homogeneous computing model; the homogeneous computing model is generated based on at least one type of memristor; obtains a speech digit recognition task, and executes the speech digit recognition task through a hybrid computing model; the hybrid computing model is generated based on at least two types of memristors.

[0111] The homogeneous computing model can refer to a reservoir computing model generated based on a same type of memristor. The hybrid computing model can refer to a reservoir computing model generated based on different types of memristors. The short-term memory (STM) task is a memory recall task in which the reservoir processes the original time series into a format from which the input value at some time delay in the past can be reconstructed. The parity check (PC) task aims to reconstruct the result of a binary parity check operation (e.g., an addition operation) on previous inputs up to some delay in the past. The speech digit recognition task can refer to a recognition task related to speech being a digit. The parity check (PC) task, for example, The memory capacity (MC PC ) is calculated according to the formula:

[0112]

[0113] In this embodiment, the target boundary time is obtained by sampling the decay boundary time of the time node of the first node number based on the second target probability density function, and the second interval duration supports dynamic adjustment; other time nodes with a second interval duration less than the target boundary time are selected as adjacent nodes of the time node, the time node and the adjacent nodes of the time node are connected to obtain a new reserve pool layer; the new reserve pool layer generates a reserve pool calculation model, and the adjustment of the network structure is realized by supporting the dynamic adjustment of the second interval duration, the reconfigurability of the generated reserve pool calculation model at the level of the memristor is realized, and the flexibility of the network structure is effectively improved. The target boundary time is obtained by sampling the second target probability density function, and the adjacent nodes determined according to the obtained target boundary time are connected, so that the flexibility of the connection between nodes is further improved, and the network computing efficiency is ensured.

[0114] As an example, the embodiment is as follows:

[0115] I. Analysis of dynamic memristor

[0116] Connectionist artificial intelligence inspired by brain neural connections (such as deep learning) has achieved unprecedented success in the past decade. The goal of artificial neural network design and optimization gradually transitions from training connection weights to designing network structure. However, behind this trend, current artificial intelligence hardware (such as GPU, FPGA, ASIC, etc.) is difficult to balance computing efficiency and network structure flexibility. Therefore, it is very important to design a network generation and calculation method that can balance computing efficiency and network structure flexibility through software and hardware collaboration. Here, the scheme considers using a single dynamic memristor to construct a complex network by using multiplexing.

[0117] Figure 4 An embodiment of the electrical characteristics of the memristor is shown in the schematic diagram; as shown in Figure 4 a figure: the evolution of the current through the device under the read voltage (3V) after the stop of the stimulation voltage pulse (5V, 1ms). Inset: magnified view of the current evolution at short time intervals before, during and after the application of the stimulation pulse. After the voltage is reduced from 5V to 3V, the current immediately drops from the peak I - to I + . b figure: spontaneous decay of the current under the read voltage from I + to the baseline steady-state value after the stimulation pulse stops. In 1000 measurements, the decay duration τ is 342ms (τ min ) and 1089ms (τ maxThe values ​​vary greatly between τ and I-. Figure c: Statistical analysis of the correlation between I+ and I- for more than 10 sets of multi-pulse (1–10) measurements, each set containing 1000 independent measurements. Figure d: Statistical analysis of the correlation between τ and I- for more than 10 sets of multi-pulse (1–10) measurements, each set containing 1000 independent measurements.

[0118] First, to assess the degree of cycle-to-cycle (C2C) variation of the device, one thousand identical and independent pulse measurements were performed on the device, and its dynamics were statistically analyzed. For example... Figure 4 As shown in Figure b, the current I is read. + The spontaneous decay time (τ) is 342 ms (τ) min ) and 1089ms(τ max The τ probability distribution of C2C looks like a Gaussian distribution truncated on both sides, where the random variable τ varies greatly between the two sides. max =1089ms) and below (τ) min Both τ and I are bounded (=342ms). + It is also related, meaning that the change in τ may originate from I. + The change in τ. When considering whether the same distribution obtained from single-pulse measurements is also reasonably applicable to the τ change of C2C measured under arbitrary pulse protocols, volatile memristors with finite τ can exhibit paired pulse promotion (or short-term promotion), i.e., I - It increases with each arriving pulse, as long as the pulse interval is shorter than τ. This can be understood as due to the time coupling between the switching events of the induced resistance of adjacent pulses. Given this, τ obtained from the last pulse in a pulse sequence or multi-pulse measurement may or may not follow the same truncated Gaussian probability distribution as obtained from a single-pulse measurement, depending on I. + And τ, is it also related to I? - Related. To address this issue, multiple sets of multi-pulse measurements were performed, each with a different number (2–10) of pulses and comprising one thousand independent experiments, recording I at the end of the last pulse. - and I + And this is τ, which ends as the last pulse. The interval between consecutive pulses is set to 200 milliseconds, which is shorter than the minimum τ recorded in single-pulse measurements. This ensures that the consecutive resistance switching events are time-coupled. Figure 4 As shown in figures c and d, although I... - The increase in the number of pulses is statistically significant, but I + and τ and I -No clear correlation is observed. This observation indicates that the memristive changes in the device ion or electron configuration caused by multiple pulses are negligible under the experimental protocol. - and I + The difference in sensitivity to configuration changes between I and I + (and τ) and I - The τ probability distribution of the same C2C obtained from single-pulse measurements is also reasonably applicable to the measurements under these multiple pulses, given the non-correlation between I Figure 5 Statistical analysis of the decay time of the memristor in one embodiment; as shown in a graph in Figure 5 a Probability distribution obtained in 10 sets of multiple-pulse (1-10) measurements, each containing 1000 independent measurements (same as in Figure 4 d of FIG. 6). This is clearly embodied in the well-overlapped distribution functions that appear in the statistical measurements of each experimental set.

[0119] As before, a dynamic memristor (such as a Ta / Ta2O x / HfO2 / Pd memristor) can temporarily couple its state at the current time (or period discretized and abstracted from the actual continuous physical time) to its state at a previous time (period). In the context of network formation, the temporal coupling between any two periods is referred to as a connection between two virtual nodes that appear in a sequential manner in the time domain. Thus, a single memristor is a time-division multiplexing unit that can be used sequentially and repeatedly. To establish a connection between two adjacent virtual nodes, the interval physical time θ (i.e., the interval duration) between two consecutive periods must be less than τ; otherwise, the two virtual nodes are independent in time and are not considered to be connected. Thus, θ becomes a key tuning factor that regulates the connectivity of the network: for a given τ, the smaller θ is, the denser the connectivity is, because the coupling time range (τ) of the virtual nodes will cover more subsequent couplings.

[0120] Although it is possible to fold a network in the spatial domain into the time domain by multiplexing dynamic memristors (for a given time slot θ), the resulting network has only trivial topological features if τ is fixed: the resulting network is simply regular. Quite to the contrary, real-world networks often have features that do not appear in simple networks (such as regular or completely random networks). In this regard, the memristor provides a source of randomness in τ, as described by the truncated Gaussian C2C probability distribution, to guarantee the topological non-triviality of the generated network.

[0121] As Figure 5Figure b in the diagram illustrates the probability distribution of τ obtained from a single-pulse measurement by fitting a truncated Gaussian function. This diagram shows that when creating a virtual node, longer / shorter τ samples from the distribution can form connections with subsequent virtual nodes that are temporally further / more distant. Since τ varies from a C2C distribution following a truncated Gaussian, the connection probability between nodes can be adjusted by θ. Specifically, virtual nodes are less than τ in physical time. min The probability of a node connecting to subsequent nodes is 1; in other words, a node must connect to D. min There are 10 subsequent nodes, of which D min =τ min / θ. The virtual node is separated from the physical time by more than τ. max The probability of a node connecting to its subsequent nodes is 0; in other words, it can never connect to the Dth node. max There are nodes and their subsequent nodes, where D is a node. max =τ max / θ.

[0122] II. Specific Implementation Steps of this Solution

[0123] Based on the characteristics of dynamic memristors described above, this embodiment proposes a memristor-inspired "probabilistic border and all-or-none connection" (PBAONC) complex network model (target complex network), the specific idea of ​​which is as follows.

[0124] Basically, there are two methods to generate complex networks with nontrivial topological characteristics: one is to modify the connections between nodes in the original rule-based network, and the other is to generate connections from scratch. The Ta / Ta2O network exhibits inherent variability as observed experimentally. x Inspired by the dynamic behavior of / HfO2 / Pd memristors, a probabilistic border and all-or-none (PBAONC) connection generation mechanism is proposed for creating complex networks. Starting with an open ring lattice with N nodes, a complex network is created under the PBAONC mechanism, where each node forms connections with its clockwise neighbors in an all-or-none (AON) manner. Specifically, a node's clockwise neighbors are categorized as near or far neighbors based on their distance from the node under consideration (measured clockwise). Each node is connected to all its near neighbors but has no connection with far neighbors (i.e., AON). For each node, the boundary between its near and far neighbors is probabilistically determined. Specifically, based on experimentally characterized Ta / Ta2O... xResistance decay time length distribution of Hf02 / Pd memristor Figure 2 where D is the distance between node i and the boundary i is sampled from the probability distribution of formula (1) in the previous section, after sampling, a specific constraint (node connection constraint) is imposed, i.e., the total number of clockwise neighbors of node i is N-i, where N is the total number of nodes on the open loop lattice (i starts from 1 at the open loop clockwise end and increases in the clockwise direction). This means that if the gap of the open loop is in the clockwise lattice path from i to j, then the clockwise node j relative to node i is not considered as a clockwise neighbor of node i. In this case, even if the distance between them (measured in the clockwise direction from node i) is less than the sampled D i , node j is not projected from node i to node j. The basic principle behind imposing this constraint is the time causality, i.e., the virtual nodes of the memristor generated in time sequence should not affect the early nodes. In order to avoid the appearance of isolated nodes, D min is also set to be non-zero, Figure 6 is a schematic diagram of a complex network in an embodiment; as Figure 6 shown, a schematic topological structure of a PBAONC network defined on the parameter space (left). Parameterized as D max =N-1=9 and D min =2. The schematic diagram of an N-node PBAONC complex network (N=10) (right). The connections are shown in thick lines. The connection range of node i in the clockwise direction is schematically shown as a thin solid concentric circular arc starting from the radial line through node i and covering other D i nodes. For nodes 9 and 10, the part / whole of the schematic arc in the clockwise direction through the radial line of the open loop lattice gap counterclockwise end node (i.e., node 10) is shown in dashed line, indicating that no connection is formed between node 9 / 10 and the nodes covered by the corresponding dashed arc.

[0125] The PBAONC mechanism can generate a series of complex networks with different edge numbers, as Figure 6 shown. Since the main concern is the complex network for information processing, two relevant metrics are considered: one is the clustering coefficient (C*), a measure of information isolation (i.e., the degree to which the network is organized into local specialized areas), and the other is the characteristic path length (L*), a measure of information integration (i.e., the ability to quickly combine pieces of specialized information from distributed areas). Contour plots of C* and L* of the networks (N=100) generated by the PBAONC mechanism as functions of D min and D max . Figure 7 is a schematic diagram of network quantization indicators in an embodiment; Figure 7L*(a) and C*(b) of PBAONC network models with 100 nodes as a function of D max and D min contours of the function. Figure 7 L* and C* along E = 1200 (c), E = 3000 (d), and E = 4800 (e) contours for PBAONC network models and E-R random network models. f plot S* contour plot as a function of PBAONC network models with 100 nodes.

[0126] As shown in a and b plots of Figure 7 contours of E (expected number of edges) are plotted. It can be seen that these contours are almost, but not strictly, perpendicular to the right diagonal D max = D min along which the networks are parameterized to form regular connections. The upper right end of the diagonal corresponds to a fully connected (FC) network, which is often implemented using reservoir computing (RC) with memristive virtual nodes. It can be clearly observed that the gradients of C* and L* in the direction of the right diagonal (E monotonically varies) are monotonic but opposite. Specifically, sparsely connected networks (lower left region) have long L* and small C*, while densely connected networks (upper right region) are the opposite. The evolution of C* and L* along a given E contour is visualized more clearly using scatter plots. As shown in c and e plots of Figure 7 At E = 1200 (sparsely connected), L* decreases along the contour between D max and D min (D max and D min monotonically but oppositely) as the disparity increases. As the connectivity becomes denser (E = 3000), the variations of C* and L* along the E contours become increasingly flat. With further increase in connection density (E = 4800), the variations of C* and L* along the E contours, while still negligible, show opposite trends compared to the sparsely connected network (E = 1200), respectively.

[0127] Many complex systems, such as brain networks, social networks, and the Internet, evolve towards an economic trade-off between minimizing wiring cost and maximizing efficiency, characterized by small L* and large C*. This ubiquitous topological feature is known as small-worldness. It also explains the optimal balance of functional segregation and integration in brain networks. To quantify the small-worldness of the entire space parameterized networks, the metric S* proposed by Humphries and Gurney is adopted, which is based on measuring the trade-off between high local clustering and short path length. As shown in Figure 7As shown in Figure f in the drawings, the gradient of S* is also most pronounced in the right diagonal direction. S* decays with increasing E when the number of nodes is fixed. Along the contour of E, there is a unique point where S* is maximized. For sparse connections (E = 600, 1200), this point lies at the maximum end of the contour of {D max -D min} while for more dense connections (E = 1800, 2400, 3000, 3600, 4200), it lies between the two ends.

[0128] In addition to C*, L* and S*, other metrics were used to quantify the properties of complex networks and compare them to other canonical complex networks, including Watts-Strogatz (W-S) small-world (SW) networks, Renyi (E-R) random networks and Barabasi-Albert (B-A) scale-free networks, under the same number of nodes (100) and the same E, as shown in Figure 8As a common reference, the FC network generated by the PBAONC mechanism is also included in each comparison group. It can be seen that the FC network generally has the smallest L* and the largest C*. Its small-world degree is very low, comparable to the E-R random network and the B-A scale-free network (E varies from 600 to 4800). Because the FC network itself is a clique, i.e., it has the largest "maximal clique size" in any of which there are two nodes. Next, the PBAONC networks with different E in each comparison group always have the second largest "maximal clique size". This can be understood as a result of the specific wiring rule that generates the PBAONC network, i.e., any one node is connected to all its near-neighbor nodes. For E = 600, the PBAONC network has the largest C* among all complex networks, and accordingly, the largest "local efficiency" (indicating the efficiency of information integration between the immediate neighbors of individual nodes). However, its L* and the corresponding "radius" are still larger than those of the W-S SW network (the other two complex networks are known to have smaller L*), thus resulting in the second highest small-world degree to the W-S SW network. As E increases, the differences in C*, L*, and S* between the PBAONC network and the W-S SW network become smaller and smaller, and the characteristics of all complex networks tend to be more and more similar to those of the FC network, as expected. Nevertheless, the first two complex networks still have the highest degree of small-worldness. Measuring the efficiency of long-range information transmission, the "global efficiency" of the PBAONC network also improves as E increases. The tree width measures the similarity between a graph and a tree. The tree width of the PBAONC network is significantly smaller than that of any other network, so it is the most tree-like network, which indicates that it is more likely to exhibit a medium-scale structure composed of small and dense parts, which represent sparsely interconnected clusters. Again, this reflects its specific wiring behavior, i.e., any one node is connected to all its near-neighbor nodes (helping to form clusters), but has no connection with any of its far-neighbor nodes (helping to form sparse interconnections between clusters).

[0129] Figure 8 Figure for one embodiment comparing the performance of the target complex network with typical complex networks; Radar chart of nine network topology indicators for PBAONC FC network, E-R random network, W-S SW network, PBAONC complex network, and B-A scale-free network with different connection densities (E = 1200 (left), E = 3000 (middle), and E = 4800 (right)). The nine indicators are clustering coefficient (C*), global efficiency (GE), local efficiency (LE), SW coefficient (S*), tree width (T), maximal clique size (MCS), radius (R), characteristic path length (L*), and wiring cost (E), respectively.

[0130] III. Applications of PBAONC complex networks

[0131] As before, the brain is a powerful computing machine that uses extremely complex neural networks. Although a complete causal mechanistic explanation of how cognitive functions arise from complex connectivity is still beyond the reach of today's neurotechnologies, connectionist models inspired by brain networks have clearly shown that complex networks can indeed efficiently perform computations.

[0132] Here, the new reservoir layers composed of PBAONC complex networks and implemented in dynamic memristors with intrinsic variability will be shown and demonstrated to be superior to previous memristor FC network reservoir layers. From the discussion in the last two sections, it is clear that Ta / Ta2O x / HfO2 / Pd memristor multiplexed N times and with a fixed time slot Θ yields various types of PBAONC networks: if NxΘ < τ min , the FC network works because even the most temporally distant virtual nodes, the first and the last, are coupled together; if Θ > τ max , there are only isolated virtual nodes because even the directly adjacent nodes are decoupled; if τ min < Θ < τ max , isolated virtual nodes can still exist. The case of Θ > τ min is beyond the current scope of interest. If Θ < τ min and NxΘ > τ max , each node is coupled in time to the subsequent appearing partial nodes, and the coupling extends only to the near nodes in time. With the appearance of new virtual nodes in each multiplexing cycle, the respective time boundaries between their near and far neighbor nodes are sampled from the τ distribution. In this way, a PBANOC complex network, Figure 9 is a schematic diagram of the new reservoir layers in one embodiment; as Figure 9 a diagram: a set of PBAONC complex network reservoir layers based on RC time-multiplexed dynamic memristors. For a single-component reservoir, a Q-dimensional input vector at a certain time is multiplied by a random NxQ mask matrix, which is converted into an N-dimensional vector representing a time input stream within an N x Θ interval. This time input stream is then fed into a dynamic memristor; that is, the dynamic memristor is time-division multiplexed with a time slot Θ, which is reused N times. Each device reservoir layer can have a different time slot allocation, and the transient dynamic responses of each memristor in the same multiplexing cycle are aligned with each other in software. The states of all virtual nodes are linearly weighted and summed by an output weight matrix W out to obtain the output of the RC system. As Figure 9 b diagram: the spectral radius of the weight matrix of a PBAONC complex network as a function of N and D maxContour plot of the function. A random weight value between -0.5 and +0.5 was assigned to each connection. The spectral radius of the weight matrix of the cPBAONC FC network and the PBAONC complex network (D max = 8) as a function of N.

[0133] An ideal reservoir should exhibit fading memory, that is, the influence of previous reservoir states on future states should gradually disappear over time. In practice, this property can be ensured if the reservoir layer weight matrix W is scaled such that its spectral radius p(W) (i.e., the largest absolute eigenvalue) satisfies p(W) < 1. Theoretical analysis also shows that if p(W) is close to 1, the reservoir layer has an optimal activity state. A random weight value between -0.5 and +0.5 was assigned to each connection in the reservoir, and p(W) was calculated. The contour plot (b-plot in Figure 9 ) shows p(W) of the PBAONC complex network reservoir layer as a function of D max and N. It can be seen that as the number of nodes increases, the optimal value of D max at which p(W) is closest to 1 decreases, while for D max = 8, the range of N (20-30) that the reservoir can have is relatively wider, and their p(W) is close to 1. From the c-plot in Figure 9 , the superiority of the PBAONC complex network reservoir layer for D max = 8 relative to the PBAONC FC (N x θ<τ min ) network reservoir layer is clear: regardless of the number of nodes, the PBAONC complex network reservoir keeps D max on average and is close to 1, while the p(W) of the PBAONC FC network reservoir is greater than 1 and increases with the number of nodes.

[0134] High dimensionality of reservoir state and its fading memory are crucial for processing time series. Here, the time series information processing capability of the PBAONC complex network reservoir was tested in short-term memory (STM) tasks, parity check (PC) tasks, and speech digit recognition tasks.

[0135] For the STM task, a binary time series input with random "0" or "1" components is used in each time step. At any time step, the respective series component is multiplied by a randomly generated (fixed throughout the processing task) binary mask matrix of size N x 1 (functionally equivalent to a synaptic weight matrix), where N is the number of reservoir nodes, resulting in a new N-dimensional vector. By time-division multiplexing, each virtual node uses the respective vector component of the N-dimensional vector for updating. At the end of each time step, all virtual nodes have been updated and the reservoir layer reaches a new state, ready to process the input in the next time step. Experimentally, the "0" and "1" vectors of the n-dimensional vector signal are represented by 1 ms voltage pulses of intensity 3 V and 5 V, respectively, with an interval of θ (also called multiplexing period duration) between consecutive pulses. Thus, the time series device will retain a duration of N x θ, after which the device in the next time step will be processed by the reservoir layer. As in the previous section, the number of multiplexing periods N (i.e. the number of virtual nodes) and the multiplexing period duration θ are varied. The transient dynamic response of the reservoir is read out by an output layer (implemented in software) that is a linear weighted sum of the reservoir node states (i.e. I - ) at the end of the processing time. One major advantage of the RC model is the fast training, since only the weights in the linear readout layer need to be trained, while the connections in the reservoir layer remain fixed. The memory capacity (MC STM ) can be quantified by the sum of the squares of the correlation of the output y k (t) and the delayed input u(t-k) over all delays, as follows:

[0136]

[0137] In addition to the fading memory property, the nonlinear dynamics of the reservoir layer are important, as they allow for linear separability of different inputs, which can be assessed using the PC task. The PC task aims to reconstruct the result of a binary parity check operation (e.g. an addition operation) on previous inputs, up to some delay in the past (e.g., 5 as the target output). The memory capacity (MC PC ) is calculated according to the following formula:

[0138]

[0139] For the PBAONC complex network reservoir layer, the contour plots (a and b plots in Figure 7 ) show the MC STM and MC PC vs. the ratio of the reservoir layer composed of PBAONC FC networks as a function of the number of virtual nodes (N) and D max (τ max / θ). Large MC is mainly in D max= 8 and N = 20-30, which is consistent with the condition to achieve the closest-to-1 p(W) Figure 9 This finding is consistent with the understanding that the reservoir layer with p(W) close to 1 has the best activity state.

[0140] To scale up the reservoir layer size or simply generate a set of reservoir layers, multiple devices can be used according to the device-to-device (D2D) variation, where the reservoir layer state is represented by the collective state of all devices. For this purpose, multiple Ta / Ta2O x / HfO2 / Pd dynamic memristors fabricated on the same substrate are used to implement a total of 600 virtual nodes. Each device is reused the same number of times (i.e., the same number of multiplexing periods) within one time step, thus generating the same number of virtual nodes. The multiplexing period duration Θ is the same for all devices. Figure 10 Related schematic diagrams for the reserve pool calculation model in one embodiment; MC STM (a) and MC PC (b) of a single PBAONC complex network reservoir layer implemented with dynamic memristors. STM (c) and MC PC (d) is a simple multi-device reservoir set (each device reservoir has the same time slot allocation), containing a total of 600 virtual nodes (each device reservoir contains N virtual nodes). Comparison of MC STM (e) and MC PC (f) of four hybrid reservoir layer groups, each containing a total of 600 virtual nodes, but with different numbers of constituent reservoir layers (thus different N). In one group, the time slot allocation (Θ) of each constituent reservoir layer is not exactly the same (thus D max g. The cochleagram of 64 frequency channels of the spoken digit "nine" from a female speaker, pre-processed by the Lyon passive ear model (left). The binary pulse sequence converted from the cochleagram by setting the spike-triggering threshold to 0.5 (right). h. The confusion matrix showing the comparison of the classification results obtained from the experimental hybrid PBAONC complex network reservoir with the correct output.

[0141] As Figure 10 c and d figures in show the measured MC STM and MC PC ratios of the reservoir layer composed of PBAONC FC networks, as a function of the number of virtual nodes generated per device (N) and D STM and MC PC contour plots, respectively. max (τ max / θ). As expected, the multi-device reservoir ensemble has improved MC compared to the single-device reservoir due to D2D variations. By using not only multiple devices but also different network parameters (D max ) for each generated reservoir layer, the RC performance can be further improved. Four such hybrid reservoir ensembles were investigated, each with a total of 600 nodes and containing several of the best-performing single PBAONC complex network reservoirs. As shown in e and f plots in Figure 10 , these PBAONC complex network reservoir ensembles achieved greater MC STM and MC PC than simple multi-device reservoir ensembles, with the reservoir ensembles parameterized to 25 devices, each device in the reservoir having 24 virtual nodes (r x N = 25 x 24 = 600) with the largest MC, where r is the number of parallel memristors. This best-performing hybrid reservoir ensemble was used in the following tasks.

[0142] Finally, a standard speech recognition benchmark was performed - isolated speech digit recognition using the hybrid PBAONC complex network reservoir ensemble. The input to the reservoir was the 64-channel sound waveform of isolated spoken digits (English 0-9) from the NIST TI46 database. Out of 500 audio samples in the TI-46 database, 450 were selected for training, with the remaining 50 samples used for testing. The RC system was implemented using 64 devices. The input signal from each individual channel was binarized into a 36-step 0 / 1 time series, Figure 10 as shown in g plot in Figure 10 . The level component in each time step was multiplied by a randomly generated binary mask matrix of size 24 x 1, represented by a 3V or 5V pulse train (1 ms in duration). Although the θ (or pulse interval) of each device in the hybrid reservoir ensemble can be different, their transient dynamic responses in the same multi-cycle were aligned with each other in software for further processing. Figure 10 h plot in Figure 10 shows the confusion matrix obtained experimentally during testing. Overall, up to 99.5% recognition rate can be achieved in the hybrid PBAONC complex network reservoir ensemble.

[0143] It should be understood that although each step in the flowchart involved in the above embodiments is shown in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in the above embodiments can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or steps or stages in other steps.

[0144] Based on the same inventive concept, the embodiments of the present application also provide a complex network generation device for implementing the above-mentioned complex network generation method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more complex network generation device embodiments provided below can refer to the limitations of the complex network generation method in the above text, which will not be repeated here.

[0145] In one embodiment, as shown in Figure 11 a complex network generation device is provided, comprising: a first determination module 1102, a second determination module 1104, a sampling module 1106, a selection module 1108 and a connection module 1110, wherein:

[0146] The first determination module 1102 is configured to determine the node arrangement structure, the node distance between the space nodes and the node connection constraint; the space nodes include nodes at different positions;

[0147] The second determination module 1104 is configured to determine the distance lower limit value and the distance upper limit value based on the decay time length of the memristor and the first interval time length; the first interval time length supports dynamic adjustment;

[0148] The sampling module 1106 is configured to sample the boundary distance of the space node based on the first target probability density function, to obtain a target boundary distance; the target boundary distance is located between the distance lower limit value and the distance upper limit value;

[0149] The selection module 1108 is configured to select other space nodes with a node distance less than the target boundary distance from the space nodes as adjacent nodes of the space nodes;

[0150] The connection module 1108 is configured to connect the space nodes and the adjacent nodes of the space nodes according to the node connection constraint under the node arrangement structure, to obtain a target complex network.

[0151] In an embodiment, the node arrangement structure comprises a circular ring arrangement structure; the first determining module 1102 is further configured to obtain the circular ring arrangement structure; after arranging the spatial nodes according to the circular ring arrangement structure, the node distance between the spatial nodes is determined in a clockwise direction; the clockwise connection rule and the open loop connection rule are obtained; and the clockwise connection rule and the open loop connection rule are taken as the node connection constraint.

[0152] In an embodiment, the selecting module 1108 is further configured to determine the node distance between the spatial node and other spatial nodes in a clockwise direction; when the node distance is less than the target boundary distance, the other spatial node corresponding to the node distance is taken as the adjacent node of the spatial node; when the node distance is not less than the target boundary distance, the other spatial node corresponding to the distance is taken as the remote node of the spatial node; and the remote node is not connected with the node.

[0153] In an embodiment, the connecting module 1110 is further configured to connect the spatial node and the adjacent node of the spatial node according to the node connection constraint under the node arrangement structure, to obtain an initial complex network; obtain a network quantization index; the network quantization index comprises a clustering coefficient, a characteristic path length and a small world coefficient; and adjust the initial complex network according to the network quantization index, to obtain a target complex network.

[0154] In an embodiment, the sampling module 1106 is further configured to sample the decay boundary time of the time node of the first node number based on the second target probability density function, to obtain a target boundary time, if the second interval time length of the memristor is not greater than the lower limit value of the decay time length, and the product of the first node number and the second interval time length is not less than the upper limit value of the decay time length; the time node comprises nodes at different time points; the target boundary time is located between the lower limit value of the decay time length and the upper limit value of the decay time length; and the second interval time length supports dynamic adjustment; the selecting module is further configured to select other time nodes with a second interval time length less than the target boundary time from the time nodes, as adjacent nodes of the time node; and the connecting module is further configured to connect the time node and the adjacent node of the time node, to obtain a new reserve pool layer; and the new reserve pool layer generates a reserve pool calculation model.

[0155] In an embodiment, as shown in FIG. 11, Figure 12 the complex network generation apparatus further comprises a processing module 1112 and an executing module 1114, wherein:

[0156] The processing module 1112 is configured to obtain an initial probability density function; determine an expected value and a variance based on a distance lower limit value and a distance upper limit value; substitute a coefficient parameter, the variance and the expected value into the initial probability density function, to obtain a probability density function; generate a key constant based on the probability density function, the distance lower limit value and the distance upper limit value; and combine the probability density function and the key constant to obtain a first target probability density function.

[0157] The execution module 1114 is further configured to obtain a short-term memory task or a parity check task, execute the short-term memory task or the parity check task through a same-type computing model, generate the same-type computing model based on at least one same-type memristor, obtain a speech digit recognition task, and execute the speech digit recognition task through a hybrid computing model, and generate the hybrid computing model based on at least two types of memristors.

[0158] The above embodiment determines the node arrangement structure, the node distance between the space nodes, and the node connection constraint; the space nodes include nodes at different positions; the distance lower limit value and the distance upper limit value are determined based on the decay duration of the memristor and the first interval duration; the first interval duration supports dynamic adjustment; the network structure is adjusted by dynamically adjusting the first interval duration, the reconfigurability of the generated complex network at the level of the memristor is achieved, and the flexibility of the network structure is effectively improved. Moreover, the target boundary distance is obtained by sampling the boundary distance of the space node based on the first target probability density function; the target boundary distance is located between the distance lower limit value and the distance upper limit value; other space nodes with a node distance less than the target boundary distance are selected as adjacent nodes of the space node; the space node and the adjacent nodes of the space node are connected according to the node connection constraint under the node arrangement structure, to obtain a target complex network; the target boundary distance is obtained by sampling through the first target probability density function, and the adjacent nodes determined according to the obtained target boundary distance are connected, the flexibility of the connection between the nodes is achieved, and the flexibility of the network structure is further improved while ensuring the network computing efficiency.

[0159] Each module in the above complex network generation device can be realized by software, hardware, and a combination thereof, in whole or in part. Each module can be embedded in or independent of a processor in a computer device in a hardware form, or can be stored in a memory in a computer device in a software form, so as to be called and executed by a processor to perform the operations corresponding to each module.

[0160] In one embodiment, a computer device is provided, which can be a terminal or a server. Taking the terminal as an example, the internal structure diagram of the terminal can be as shown in FIG. 8. Figure 13The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to perform wired or wireless communication with external terminals. The wireless communication can be achieved through WIFI, mobile cellular network, NFC (Near Field Communication) or other technologies. The computer program is executed by the processor to implement a complex network generation method. The display unit of the computer device is configured to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0161] Those skilled in the art can understand that, Figure 13 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0162] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-mentioned embodiments.

[0163] In one embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program is executed by a processor to implement the above-mentioned embodiments.

[0164] In one embodiment, a computer program product is provided, including a computer program, and the computer program is executed by a processor to implement the above-mentioned embodiments.

[0165] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0166] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto. The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, not all possible combinations of the technical features in the above-mentioned embodiments are described, however, as long as the combination of these technical features does not exist contradictory, it should be considered as the scope of the present application.

[0167] The above embodiments only express several implementation ways of the present application, and the description is specific and detailed, but it should not be understood as a limitation to the patent scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for generating a complex network, characterized by, The method comprises: determining a node arrangement structure, node distances between space nodes, and node connection constraints; the space nodes include nodes at different positions; determining a distance lower limit value and a distance upper limit value based on a decay time length of a memristor and a first interval time length; the first interval time length supports dynamic adjustment; sampling a boundary distance of the space node based on a first target probability density function to obtain a target boundary distance; the target boundary distance is between the distance lower limit value and the distance upper limit value; selecting other space nodes with a node distance from the space node less than the target boundary distance as adjacent nodes of the space node; connecting the space node and the adjacent nodes of the space node according to the node connection constraints under the node arrangement structure to obtain a target complex network.

2. The method of claim 1, wherein, The node arrangement structure comprises a circular ring arrangement structure; the determination of the node arrangement structure, the node distances between the space nodes, and the node connection constraints comprises: obtaining the circular ring arrangement structure; determining the node distances between the space nodes in a clockwise direction after arranging the space nodes according to the circular ring arrangement structure; obtaining a clockwise connection rule and an open loop connection rule; taking the clockwise connection rule and the open loop connection rule as the node connection constraints.

3. The method of claim 1, wherein, The method further comprises: obtaining an initial probability density function; determining an expected value and a variance based on the distance lower limit value and the distance upper limit value; substituting a coefficient parameter, the variance, and the expected value into the initial probability density function to obtain a probability density function; generating a key constant based on the probability density function, the distance lower limit value, and the distance upper limit value; combining the probability density function and the key constant to obtain a first target probability density function.

4. The method of claim 1, wherein, The selection of other space nodes with a node distance from the space node less than the target boundary distance as adjacent nodes of the space node comprises: determining the node distances between the space node and other space nodes in a clockwise direction; when the node distance is less than the target boundary distance, taking the other space node corresponding to the node distance as an adjacent node of the space node; The method further comprises: when the node distance is not less than the target boundary distance, taking the other space node corresponding to the node distance as a remote node of the space node; the remote node is not connected to the node.

5. The method of claim 1, wherein, The connection of the space node and the adjacent nodes of the space node according to the node connection constraints under the node arrangement structure to obtain a target complex network comprises: connecting the space node and the adjacent nodes of the space node according to the node connection constraints under the node arrangement structure to obtain an initial complex network; obtaining network quantization indicators; the network quantization indicators include a clustering coefficient, a characteristic path length, and a small-world coefficient; adjusting the initial complex network according to the network quantization indicators to obtain a target complex network.

6. The method of claim 1, wherein, The method further comprises: If the second interval length of the memristor is not greater than the lower limit of the decay length, and the product of the first node number and the second interval length is not less than the upper limit of the decay length, a target boundary time is obtained by sampling a decay boundary time of a time node of the first node number based on a second target probability density function, the time node includes nodes at different time points, the target boundary time is located between the lower limit of the decay length and the upper limit of the decay length, and the second interval length supports dynamic adjustment. Other time nodes with a second interval length less than the target boundary time are selected as adjacent nodes of the time node. The time node and the adjacent nodes of the time node are connected to obtain a new reserve pool layer, and the new reserve pool layer generates a reserve pool calculation model.

7. The method of claim 6, wherein, The reserve pool calculation model is a same type calculation model or a mixed calculation model, and the method further includes: A short-term memory task or a parity check task is obtained, and the short-term memory task or the parity check task is executed by the same type calculation model, and the same type calculation model is generated based on at least one same type memristor; A speech digital recognition task is obtained, and the speech digital recognition task is executed by the mixed calculation model, and the mixed calculation model is generated based on at least two types of memristors.

8. A complex network generating apparatus characterized by comprising: The device includes: A first determination module is configured to determine a node arrangement structure, node distances between space nodes, and node connection constraints, and the space nodes include nodes at different positions; A second determination module is configured to determine a distance lower limit value and a distance upper limit value based on a decay length and a first interval length of a memristor, and the first interval length supports dynamic adjustment; A sampling module is configured to sample a boundary distance of the space nodes based on a first target probability density function to obtain a target boundary distance, and the target boundary distance is located between the distance lower limit value and the distance upper limit value; A selection module is configured to select other space nodes with a node distance less than the target boundary distance as adjacent nodes of the space nodes; A connection module is configured to connect the space nodes and the adjacent nodes of the space nodes according to the node connection constraints under the node arrangement structure to obtain a target complex network. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

11. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

Images