Task processing method and system for collaborative learning of multiple intelligent nodes

Abstract

The invention provides a task processing method and system for multi-intelligent-node collaborative learning, and relates to the technical field of artificial intelligence, and the method comprises the steps: employing a plurality of intelligent nodes to construct a target distributed intelligent network with a plurality of node communities; the method comprises the steps of obtaining a domain task, determining a target intelligent node and a target domain label, establishing a task evaluation function and an increment residual layer, carrying out collaborative learning in a node community corresponding to the target domain label, and generating an increment residual module, the incremental residual module has a layer correspondence relationship with a network layer with a target domain label in the basic artificial neural network under the target intelligent node; and the target intelligent node processes the domain task based on the incremental residual module, and shares the incremental residual module to other intelligent nodes in the node community corresponding to the target domain label, so that the problems that a large amount of data needs to be centrally processed when new domain skills are added to an existing large model, and privacy leakage exists in the training process are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly to a task processing method and system for collaborative learning of multiple intelligent nodes. Background Art

[0002] Existing data-driven deep neural network technologies rely on a large amount of data. One way of learning is to gather all the data into a data center to centrally train a certain deep neural network model, which will bring data security problems, mainly manifested as the leakage of private data or confidential data. Some collaborative learning attempts to keep the data locally on intelligent nodes and only transfer the gradients during model training to the intelligent nodes participating in collaborative learning. Although this protects local data from leakage to a certain extent, multiple intelligent nodes are training the same model, and the final model obtained is homogeneous, and the complete model needs to be synchronized regularly, otherwise the transferred gradients cannot be shared among different nodes. Summary of the Invention

[0003] In order to overcome the above technical defects, the purpose of the present invention is to provide a task processing method and system for collaborative learning of multiple intelligent nodes, so as to solve the problems that existing large models need to centrally process a large amount of data to add new domain skills and there are privacy leaks in the training process.

[0004] The present invention discloses a task processing method for collaborative learning of multiple intelligent nodes, including: Constructing a target distributed intelligent network with multiple node communities by using multiple intelligent nodes, where each intelligent node includes a basic artificial neural network, and the basic artificial neural network is composed of a backbone network and an incremental residual module. Each network layer is associated with at least one domain label, and each node community corresponds to a unique domain label, so that intelligent nodes with the same domain label form a node community; Obtaining a domain task, determining a target intelligent node and a target domain label, establishing a task evaluation function and an incremental residual layer, and performing collaborative learning within the node community corresponding to the target domain label to generate an incremental residual module acting on the basic artificial neural network under the target intelligent node, where the incremental residual module has a layer correspondence relationship with the network layer with the target domain label in the basic artificial neural network under the target intelligent node; The target intelligent node processes the domain task based on the incremental residual module and shares the incremental residual module with other intelligent nodes in the node community corresponding to the target domain label.

[0005] Preferably, the establishing a task evaluation function and an incremental residual layer, and performing collaborative learning within the node community corresponding to the target domain label to generate an incremental residual module acting on the basic artificial neural network under the target intelligent node includes: Randomly initialize the parameters of the incremental residual layer, and screen out the intelligent nodes for collaborative learning within the node community; In the basic artificial neural networks of the target intelligent node and each intelligent node for collaborative learning, splice the incremental residual layer to respectively execute local tasks corresponding to the domain tasks, output a global evaluation value based on the task evaluation function, and iteratively optimize the parameters of the incremental residual layer to generate the incremental residual module.

[0006] Preferably, the target intelligent node processes the domain task based on the incremental residual module, including: Splice the incremental residual module with the basic artificial neural network under the target intelligent node according to the layer correspondence relationship, where the i-th incremental residual layer under the incremental residual module is connected in parallel to the i-th layer with the target domain label in the basic artificial neural network under the target intelligent node; Obtain the output of the (i - 1)-th layer of the basic artificial neural network under the target intelligent node, input it into the i-th layer of the basic artificial neural network and the i-th incremental residual layer, and after fusion of the output, input it into the (i + 1)-th layer of the basic artificial neural network.

[0007] Preferably, when the basic artificial neural network under the target intelligent node contains multiple domain labels: Obtain the output of the (i - 1)-th layer of the basic artificial neural network under the target intelligent node, and after processing by the network layer corresponding to other domain labels, fuse it into the output of the target intelligent node processing the domain task based on the incremental residual module.

[0008] Preferably, any incremental residual layer in the incremental residual module includes an upper backbone network label, a lower backbone network label, and a backbone network; The i-th incremental residual layer is connected to the (i + 1)-th layer of the basic artificial neural network under the target intelligent node through the upper backbone network label, and is connected to the (i - 1)-th layer of the basic artificial neural network under the target intelligent node through the lower backbone network label, and splice the incremental residual module with the basic artificial neural network under the target intelligent node.

[0009] Preferably, before establishing the incremental residual layer, it further includes: Determine the feature description of the task according to the domain task; Broadcast the feature description of the task within the node community corresponding to the target domain label; When the intelligent nodes in the node community that include the incremental residual blocks corresponding to the domain task respond, obtain the incremental residual blocks according to the sharing within the node community, and replace them to generate the incremental residual module acting on the basic artificial neural network under the target intelligent node.

[0010] Preferably, after sharing the incremental residual module to other intelligent nodes in the node community corresponding to the target domain label, the method further includes: For any one of the other intelligent nodes, obtain the new residual layer code sequence of the incremental residual module to self-check whether there is a missing incremental residual layer; If so, request other intelligent nodes in the node community to share the incremental residual module.

[0011] Preferably, after sharing the incremental residual module to other intelligent nodes in the node community corresponding to the target domain label, the method includes periodically check whether there are conflicting residual layers in each intelligent node in the node community, where the conflicting residual layers are different incremental residual layers generated for the same domain task; If so, use a task evaluation function to determine the optimal incremental residual layer.

[0012] The present invention also provides a task processing system for multi-intelligent node collaborative learning, which executes the above task processing method, including: a plurality of intelligent nodes, constructing a target distributed intelligent network with a plurality of node communities, where each intelligent node includes: a basic artificial neural network, where each network layer in the basic artificial neural network is associated with at least one domain label, and each node community corresponds to a unique domain label, so that intelligent nodes with the same domain label form a node community; an incremental residual module, which is used to perform collaborative learning generation within the node community corresponding to the target domain label of the domain task according to the obtained domain task, where the incremental residual module has a layer correspondence relationship with the network layer with the target domain label in the basic artificial neural network under the target intelligent node; The incremental residual module is shared to other intelligent nodes in the node community corresponding to the target domain label.

[0013] Preferably, the basic artificial neural network of each intelligent node adopts a Transformer structure, and the incremental residual block uses an adapter module; Each intelligent node located in the same node community shares the network layer structure corresponding to the domain label.

[0014] After adopting the above technical solution, compared with the prior art, the following beneficial effects are achieved: The present application provides a task processing method and system for collaborative learning of multiple intelligent nodes. The intelligent nodes are divided into different node communities according to domain tags, and incremental learning is implemented within each node community using incremental residual blocks. By transmitting the structure and weight parameters of the newly added residual blocks to node community collaboration, the fusion of the newly added residual blocks for the same task can be achieved, solving the problems that existing large models need to centrally process a large amount of data to add new domain skills and there is privacy leakage in the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 FIG. is a flowchart of the first embodiment of the task processing method and system for collaborative learning of multiple intelligent nodes according to the present invention; Figure 2 FIG. is a schematic diagram showing node communities in the first and second embodiments of the task processing method and system for collaborative learning of multiple intelligent nodes according to the present invention; Figure 3 FIG. is a schematic diagram of the network structure of any intelligent node in the first and second embodiments of the task processing method and system for collaborative learning of multiple intelligent nodes according to the present invention; Figure 4 FIG. is a schematic flowchart of an intelligent node in the first and second embodiments of the task processing method and system for collaborative learning of multiple intelligent nodes according to the present invention to obtain incremental residual blocks from other intelligent nodes in the node community; Figure 5 FIG. is a schematic diagram of the network structure of the connection between the basic artificial neural network and the incremental residual module in the first and second embodiments of the task processing method and system for collaborative learning of multiple intelligent nodes according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The advantages of the present invention are further elaborated below in conjunction with the accompanying drawings and specific embodiments.

[0017] Here, the exemplary embodiments will be described in detail, and the examples are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0018] The terms used in the present disclosure are only for the purpose of describing specific embodiments and are not intended to limit the present disclosure. The singular forms "a", "the" and "said" used in the present disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0019] It should be understood that although the terms first, second, third, etc. may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this disclosure, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0020] In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms longitudinal, transverse, upper, lower, front, rear, left, right, vertical, horizontal, top, bottom, inner, outer, etc. are based on the orientation or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0021] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms installation, connection, and coupling should be understood in a broad sense. For example, it may be a mechanical connection or an electrical connection, or may also be the communication inside two elements. It may be directly connected, or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0022] In the subsequent description, the suffixes such as "module", "component", or "unit" used to represent elements are only for the convenience of the description of the present invention, and they do not have a specific meaning in themselves. Therefore, "module" and "component" can be used interchangeably.

[0023] Embodiment 1: This embodiment discloses a task processing method for multi-intelligent node collaborative learning. The task includes natural language processing and can be loaded into a computing system interconnected by multiple intelligent nodes, and each intelligent node is composed of a neural network. Without the need to transfer learning data between intelligent nodes, they can collaborate in learning. Among different intelligent nodes, the basic artificial neural networks (models) in the intelligent nodes do not need to be exactly the same, and there is no need to transfer model gradients during collaborative learning. Specifically, refer to Figures 1-5 , including: S10: Use multiple intelligent nodes to construct a target distributed intelligent network with multiple node communities. Each intelligent node includes a basic artificial neural network, and the basic artificial neural network is composed of a backbone network and an incremental residual module. Each network layer is associated with at least one domain label, and each node community corresponds to a unique domain label, so that the intelligent nodes with the same domain label form a node community; In this embodiment, the basic artificial neural networks of each intelligent node in the target distributed intelligent network can adopt the same pre-trained language model as the backbone network, and at the same time can have different incremental residual modules, so as to generate different basic artificial neural network structures. All intelligent nodes in the target distributed intelligent network are divided into multiple node communities. Each node community has and only has one domain label. There can be overlapping intelligent nodes between different node communities. Collaborative learning of multiple intelligent nodes is carried out within one community. Each intelligent node has at least one domain label, and the domain label can represent the domain knowledge possessed by the (intelligent) node. A node can contain multiple domain labels; each intelligent node in the node community must contain the same domain label as the community; an intelligent node can be located in multiple communities in the same domain; when an intelligent node contains multiple domain labels, it can be located in multiple domain communities. As an example, refer to Figure 2 , fields D1, D2, and D3 respectively form communities C1, C2, and C3. Node N7 has fields D1, D2, and D3, and node N8 has nodes D2, D3, etc.

[0024] It should be noted that the domain label / knowledge is embodied as the causal relationship between concepts and concepts, and can be represented as a multi-layer neural network structure and neural network weight parameters through the training learning process. Among them, the shallow neural network layer corresponds to the underlying concepts in the domain knowledge, and the high-level neural network layer corresponds to the high-level concepts in the domain knowledge; for node communities with non-empty intersection nodes, the multi-layer neural networks represented by the intelligent nodes in it will share the shallow neural network layer.

[0025] At the starting moment when the system applying the natural language task processing method of collaborative learning of multiple intelligent nodes provided in this embodiment is executed, it can be set that all intelligent nodes have an initialized neural network (i.e., the basic artificial neural network); and when two intelligent nodes have the same domain label, they will both contain the initialized neural network layer corresponding to the domain knowledge, which can include the backbone network and the jointly owned incremental residual module.

[0026] S20: Obtain a domain task, such as Chinese-Japanese translation, determine the target intelligent node and the target domain label, establish a task evaluation function and an incremental residual layer, and perform collaborative learning within the node community corresponding to the target domain label to generate an incremental residual module acting on the basic artificial neural network under the target intelligent node. Among them, the incremental residual module has a layer correspondence relationship with the network layer with the target domain label in the basic artificial neural network under the target intelligent node; This task is a task with domain labels. For new tasks within its own domain labels, in this embodiment, the intelligent node learns new knowledge, and the learned new knowledge is represented as an incremental residual neural network layer (i.e., an incremental residual module) parallel to the neural network layers in the basic artificial neural network.

[0027] Specifically, for establishing the task evaluation function and the incremental residual layer, collaborative learning is carried out within the node community corresponding to the target domain label to generate an incremental residual module acting on the basic artificial neural network under the target intelligent node, including: randomly initializing the parameters of the incremental residual layer, screening out the intelligent nodes for collaborative learning within the node community; splicing the incremental residual layer in the basic artificial neural networks under the target intelligent node and each intelligent node for collaborative learning to respectively execute local tasks corresponding to the domain task, outputting a global evaluation value based on the task evaluation function, and iteratively optimizing the parameters of the incremental residual layer to generate the incremental residual module.

[0028] During the above collaborative learning process, the intelligent nodes for collaborative learning execute local tasks corresponding to the domain task, which are tasks matched according to the task description features of the domain task. The description features of the task can be based on the compressed encoding output by several layers at the bottom of the basic artificial neural network (with all domain labels / shared network structures).

[0029] Specifically, as an example, perform the following step c1: When the node starts to learn the new task set of the domain , formulate a new task objective evaluation function Obj, generate a column of incremental residual layers corresponding to the existing neural network layers (i.e., the basic artificial neural network of this node) , is the highest number of layers of the neural network structure. The incremental residual layer contains two types of parameters. One type is the parameters describing the structure of the incremental residual layer itself and the parameters of the connection structure with the original network (basic artificial neural network), which are jointly denoted as τ. The other type is the parameters of the network weights in the incremental residual layer, denoted as θ; randomly initialize the above two types of parameters, and use an optimization algorithm to learn the weight parameters in the incremental residual layer to obtain the incremental residual layer and the evaluation value and the number of new tasks , and then send the incremental residual layer to all intelligent nodes in the communities that contain the node and whose domain label is .

[0030] c2: Node extracts the new task set The features, as well as the new task objective evaluation function Obj, are sent to the set of intelligent nodes in all communities that contain node and whose domain label is ; ; c3: For each node in the node set , According to the features of the new task set , select from its own task data the task data set that meets the permission requirements and matches the features of the new task set ; The node Splice the incremental residual layer into the neural network of this node according to the layer correspondence to form a temporary neural network model; The node Uses the task data set and the objective evaluation function Obj to evaluate the temporary neural network model to obtain the evaluation value of the temporary neural network, and then and the new task quantity for node evaluation are passed to the initiating node .

[0031] c4: The node Aggregates the evaluation values and evaluation task quantities from each node in , adds its own evaluation task quantity and evaluation value to form an evaluation value set and an evaluation task quantity set , and calculates the global evaluation value V by weighted average accordingly; Repeat the process from c1 to c4, record the number of repetitions, and use a gradient-free optimization algorithm to optimize the parameters τ and θ of the newly added residual layer until the global evaluation value V is greater than the set threshold or the number of repetitions exceeds the set number of times, and fix the parameters, that is, generate the incremental residual module.

[0032] Specifically, before establishing the incremental residual layer, it also includes: determining the feature description of the task according to the domain task; broadcasting the feature description of the task in the node community corresponding to the target domain label; when an intelligent node corresponding to the incremental residual block of the domain task in the node community responds, obtain the incremental residual block according to the sharing in the node community and replace it to generate an incremental residual module acting on the basic artificial neural network under the target intelligent node.

[0033] As mentioned above, this feature description is also applied in the step of generating the incremental residual module. The feature description of this task is the requirement of the corresponding incremental residual block of this task, which is the compressed encoding output by several layers at the bottom of the backbone of the neural network (basic artificial neural network). In this embodiment, if there is a node in the node community already contains new tasks for all incremental residual blocks, and the access rights of these incremental residual blocks are in a shareable state, then the data structure objects of these incremental residual blocks, that is, layer module objects, are sent to the nodes . That is, there is no need to use the above optimization algorithm to generate incremental residual modules anymore. Instead, they are directly obtained through sharing within the node community to improve data / task processing efficiency. By passing the structure and weight parameters of the newly added residual blocks to the community collaboration nodes, without sending any information such as node local data, or data backpropagation gradients, or data features, etc. to other nodes in the node community, the security is also relatively high.

[0034] In this embodiment, when an intelligent node receives a new task and this task cannot be effectively processed by the existing neural network model, a new incremental residual block needs to be generated by the incremental residual layer. The intelligent node can share the features and evaluation functions of the task with other nodes and request them to provide feedback and evaluation values, so as to obtain the incremental residual layer with optimal parameters, that is, the determined incremental residual block.

[0035] S30: The target intelligent node processes the domain task based on the incremental residual module and shares the incremental residual module with other intelligent nodes in the node community corresponding to the target domain label.

[0036] Specifically, the target intelligent node processes the domain task based on the incremental residual module. Refer to Figure 3 , including: splicing the incremental residual module with the basic artificial neural network under the target intelligent node according to the layer correspondence relationship, where the i-th incremental residual layer (Res-i) in the incremental residual module is connected in parallel to the i-th layer (layer-i) with the target domain label in the basic artificial neural network under the target intelligent node; obtaining the output of the (i - 1)-th layer of the basic artificial neural network under the target intelligent node, inputting it into the i-th layer of the basic artificial neural network and the i-th incremental residual layer, fusing the outputs and then inputting them into the (i + 1)-th layer of the basic artificial neural network.

[0037] As supplementary explanation, in this embodiment, the backbone network of the neural network (basic artificial neural network) in the intelligent node can adopt the Transformer structure, and the incremental residual block uses the Adaptor module. The structure and weight parameters of the backbone and incremental residual block of the neural network are stored in a data structure. The basic unit of this data structure is the layer module, that is, the layers in the Transformer network and the layers in the Adaptor module can both be stored using the layer module data structure; in addition to serving as the network structure parameters and network weight parameters required for the neural network Transformer Layer and Adaptor, the layer module of the Transformer network contains domain labels, and the layer module of the incremental residual block contains: the ID of this incremental residual block, domain labels (such as D1, D2, D3, etc. in the figure), task labels (such as T1, T2, etc. in the figure), and task feature encoding (i.e., the feature description of the above tasks).

[0038] Since the target distributed intelligent network in this embodiment is layer-based, when the incremental residual module is spliced with the basic module and the two themselves have a layer correspondence relationship, the i-th layer of the basic artificial neural network is made to be in parallel with the i-th layer in the incremental residual module. Specifically, any incremental residual layer in the incremental residual module includes an upper backbone network label, a lower backbone network label, and a backbone network; the i-th incremental residual layer is connected to the (i + 1)-th layer of the basic artificial neural network under the target intelligent node through the upper backbone network label, and is connected to the (i - 1)-th layer of the basic artificial neural network under the target intelligent node through the lower backbone network label, and the incremental residual module is spliced with the basic artificial neural network under the target intelligent node. That is, the above upper backbone network label and lower backbone network label are used as connection identifiers for the i-th incremental residual layer, so that the generated incremental residual layer can autonomously connect to the network layer under the corresponding domain label in the basic artificial neural network.

[0039] In this embodiment, as described above, a node can contain multiple domain labels. The above domain tasks only rely on the incremental learning under the incremental residual block corresponding to one domain label (node community). Therefore, when the basic artificial neural network under the target intelligent node contains multiple domain labels, that is, when the intelligent node is located in multiple node communities, this incremental residual block only updates the network parameters corresponding to this domain label. Specifically: obtain the output of the (i - 1)-th layer of the basic artificial neural network under the target intelligent node, and after being processed by the network layer corresponding to other domain labels, it is fused into the output of the target intelligent node for processing the domain task based on the incremental residual module. That is, the network layer corresponding to other domain labels will not be connected to this incremental residual block and will not affect the network layer corresponding to other domain labels. That is, different communities with the same domain label can achieve the fusion of the incremental residual blocks for the same task and will not affect the data processing under different domains.

[0040] In this embodiment, as a preferred embodiment, after sharing the incremental residual module to other intelligent nodes in the node community corresponding to the target domain label, it further includes: for any one of the other intelligent nodes, obtaining the new residual layer code sequence of the incremental residual module to self-check whether there is a missing incremental residual layer; if so, requesting other intelligent nodes in the node community to share the incremental residual module.

[0041] In this embodiment, freezing the structural parameter τ and weight parameter θ of the incremental residual layer can be regarded as creating a new residual layer and giving this new residual layer a code , and the information of domain id, task id and creation time is included in the encoding of this code; and record the domain at this time Except for the initial layer, the code sequences of all other pre-existing new residual layers . The node obtaining new task knowledge in the new domain can transfer the new knowledge, that is, the parameters τ and θ of the new residual layer, and the code , the code sequence of the pre-order new residual layer , to the nodes with the same domain label .

[0042] The node checks whether it already contains all the pre-order new residual layers according to the code sequence of the pre-order new residual layer ; if there are missing pre-order new residual layers, on the one hand, mark the codes of the missing pre-order new residual layers and their positions in the neural network (incremental residual module) structure; on the other hand, it can request other nodes in the community where the node itself is located and the domain label is to obtain the missing pre-order new residual layers, and other nodes decide whether to share them with the node according to whether they contain these pre-order new residual layers and the permission settings for these residual layers; In this embodiment, as a preferred embodiment, after sharing the incremental residual module to other intelligent nodes in the node community corresponding to the target domain label, it may further include: periodically checking whether there are conflict residual layers in each intelligent node in the node community, where the conflict residual layer is different incremental residual layers generated for the same domain task; if so, using a task evaluation function to determine the optimal incremental residual layer.

[0043] It is understandable that periodically, it is checked whether there are newly added residual layers with the same task ID but different codes among different nodes in the same community (which can be parsed from the above-mentioned codes of the newly added residual layers). In this embodiment, they are called conflicting residual layers. If there are any, a task set can be constructed in the node community, and the evaluation function of this task is used to evaluate and compare the neural network models containing the conflicting residual layers. The newly added residual layer with a higher evaluation value is selected from the conflicting residual layers as the winning residual layer, and other conflicting residual layers are eliminated. It can be parsed from the above). In this embodiment, if there are any, a task set can be constructed in the node community, and the evaluation function of this task is used to evaluate and compare the neural network models containing the conflicting residual layers. The newly added residual layer with a higher evaluation value is selected from the conflicting residual layers as the winning residual layer, and other conflicting residual layers are eliminated.

[0044] In the target distributed intelligent network in this embodiment, first, the intelligent nodes in the network are divided into different communities according to the domain labels. The intelligent nodes in the same community can decide whether to participate in a multi-intelligent-node collaborative learning according to their own wishes (specifically, the screening conditions of the nodes that perform collaborative learning in the node community corresponding to the target domain label can be adjusted). Secondly, during the collaborative learning process, an incremental residual module is used to implement incremental learning. The structural parameters and weight parameters of the incremental residual module both adopt a non-gradient optimization algorithm. By transmitting the structure and weight parameters of the newly added residual block to the community collaboration nodes, without sending any information such as the node local data, or the data backpropagation gradient, or the data features to other nodes in the community, it is ensured that the data and its associated objects only remain in the node local. Finally, during the collaborative learning process, different communities with the same domain label can realize the fusion of the newly added residual blocks of the same task, achieve the cognitive alignment of different node communities, and thus realize the task processing of multi-intelligent-node collaborative learning.

[0045] Embodiment 2: The present invention also provides a task processing system for multi-intelligent-node collaborative learning, which executes the task processing (natural language processing) method for multi-intelligent-node collaborative learning described in Embodiment 1 above, including: a plurality of intelligent nodes, which can be loaded on one or more computer devices, and a target distributed intelligent network with multiple node communities is constructed. Each intelligent node includes: a basic artificial neural network, where each network layer in the basic artificial neural network is associated with at least one domain label, and each node community corresponds to a unique domain label, so that the intelligent nodes with the same domain label form a node community; an incremental residual module, which is used to perform collaborative learning generation in the node community corresponding to the target domain label of the domain task according to the obtained domain task. Among them, the incremental residual module has a layer correspondence relationship with the network layer with the target domain label in the basic artificial neural network of the target intelligent node; the incremental residual module is shared with other intelligent nodes in the node community corresponding to the target domain label.

[0046] Optionally, the basic artificial neural network of each intelligent node adopts a Transformer structure, and the incremental residual block uses an adapter module; the network layer structures corresponding to the domain labels are shared by the intelligent nodes located within the same node community.

[0047] As an example, taking a computer as an intelligent node, each computer contains at least one domain label and a neural network (basic artificial neural network) corresponding to the domain label, where the structure of the neural network adopts a Transformer structure as described above, and multiple intelligent nodes are interconnected through the Internet to form an exclusive intelligent network; Multiple intelligent nodes with the same domain label logically form a node community. In an exclusive intelligent network, there are multiple domains, and there can be an overlap of multiple node communities in one domain. Different communities in the same domain can have overlapping intelligent nodes, and an intelligent node with multiple domain labels can also exist in different communities with different domain labels; Specifically, as an example, the backbone of the basic (neural) network in the intelligent node adopts a Transformer structure, and the incremental residual block of the neural network uses an Adaptor module; the backbone of this neural network is pre-trained to determine the initial parameters; different intelligent nodes containing the same domain include the same backbone of the neural network corresponding to the domain; the backbones of neural networks in different domains can share the same Transformer Layer; for example Figure 3 The white rectangular blocks constitute the backbone of the neural network in the intelligent node; multiple domain tasks with the same modal input will share a part of the neural network layers in the backbone of the neural network; when the intelligent node is initialized, the backbone of the neural network is initialized with pre-trained parameters; afterwards, when encountering new tasks, the learning of the new tasks generates the incremental residual blocks of the neural network; The structures and weight parameters of the backbone and incremental residual blocks of the neural network are stored in a data structure, and the basic unit of this data structure is the layer module. The layers in the Transformer and the layers in the Adaptor can both be stored using the layer module data structure; in addition to being the network structure parameters and network weight parameters required for the neural network Transformer Layer and Adaptor, the layer module Figure 3 and Figure 5 As shown, there are also the following special points: the layer module of the backbone network contains the domain label; the layer module of the task incremental residual block contains: the ID of this incremental residual block, the domain label, the task label, the task feature encoding; the ID of the previous task incremental residual block of this task incremental residual block.

[0048] The system provided by this embodiment can perform the following: An intelligent node Start learning a new task in a certain field When it does so, all the node communities in the field where it is located broadcast the requirement to obtain the incremental residual blocks corresponding to this task; the feature description of the task in the broadcast content is the compressed encoding of the outputs of several bottom layers of the neural network backbone; If there is a node in the node community that already contains all the incremental residual blocks of the new task and the access rights of these incremental residual blocks are in a shareable state, then the data structure objects of these incremental residual blocks, that is, the layer module objects, are sent to the node ; if it is found that the ID of the previous task incremental residual block is missing after receiving the layer module object, then initiate the learning of the missing previous task as described in Embodiment 1; If no node in the node community contains or can share the incremental residual blocks corresponding to the new task then use the neuroevolution algorithm to learn a group of Adaptor blocks as the incremental residual blocks corresponding to the new task ; when implementing the neuroevolution algorithm, in each round of the population (task set), optionally, the top 10% of the individuals (tasks) that are superior are selected and sent to several collaborative learning nodes in the community , select the collaborative learning nodes, and in addition to these, the following are also sent to these collaborative learning nodes: (1) the feature description of the task, (2) the task evaluation function; each collaborative learning node performs the following actions: (1) respectively find several tasks in the local data that match the task description features and do not exceed the number limit, (2) according to the data structure of the received incremental residual blocks, that is, the layer module object, insert the incremental residual blocks on the basis of the existing local neural network (basic artificial neural network) to obtain the verified neural network, (3) use the verified neural network to infer the selected local tasks, and the output results are processed by the evaluation function to obtain the evaluation values, (4) and then return the evaluation results to the collaborative learning initiating node ; the collaborative learning initiator merges the evaluation values of all collaborative learning nodes and its own local task evaluation value as the final evaluation value of this round of the population, and the subsequent processing is similar to that of the neuroevolution algorithm; when the neuroevolution algorithm ends, the structure and weight parameters of the incremental residual blocks are determined; For illustration, to become a collaborative learning node the following conditions need to be met: (1) the node is willing to participate in collaborative learning (preset condition); (2) the node already contains all the previous incremental residual blocks for learning this task, which can be achieved by checking whether the IDs of the previous incremental residual blocks exist in the current network; finally, the collaborative learning initiating node sends the new task Send the layer module object of the corresponding incremental residual block to the shared intelligent nodes in the community.

[0049] It should be noted that the embodiments of the present invention have better implementability and do not impose any form of limitation on the present invention. Any person skilled in the art may use the technical content disclosed above to modify or decorate it into equivalent effective embodiments. However, as long as it does not depart from the content of the technical solution of the present invention, any modification, equivalent change or decoration made to the above embodiments according to the technical essence of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A task processing method for collaborative learning of multiple intelligent nodes, characterized in that Including: Construct a target distributed intelligent network with multiple node communities by using multiple intelligent nodes. Each intelligent node includes a basic artificial neural network, and the basic artificial neural network is composed of a backbone network and an incremental residual module. Each network layer is associated with at least one domain label, and each node community corresponds to a unique domain label, so that intelligent nodes with the same domain label form a node community; Obtain a domain task, determine the target intelligent node and the target domain label, establish a task evaluation function and an incremental residual layer, and perform collaborative learning within the node community corresponding to the target domain label to generate an incremental residual module acting on the basic artificial neural network under the target intelligent node. Among them, the incremental residual module has a layer correspondence relationship with the network layer with the target domain label in the basic artificial neural network under the target intelligent node; The target intelligent node processes the domain task based on the incremental residual module and shares the incremental residual module with other intelligent nodes in the node community corresponding to the target domain label.

2. The task processing method according to claim 1, wherein The establishment of the task evaluation function and the incremental residual layer, and the collaborative learning within the node community corresponding to the target domain label to generate an incremental residual module acting on the basic artificial neural network under the target intelligent node includes: Randomly initialize the parameters of the incremental residual layer and select the intelligent nodes for collaborative learning within the node community; Splice the incremental residual layer in the basic artificial neural networks under the target intelligent node and each intelligent node for collaborative learning to respectively execute local tasks corresponding to the domain task, output a global evaluation value based on the task evaluation function, and iteratively optimize the parameters of the incremental residual layer to generate the incremental residual module.

3. The task processing method according to claim 1, wherein The target intelligent node processes the domain task based on the incremental residual module, including: Splice the incremental residual module with the basic artificial neural network under the target intelligent node according to the layer correspondence relationship, where the i-th incremental residual layer under the incremental residual module is connected in parallel to the i-th layer with the target domain label in the basic artificial neural network under the target intelligent node; Obtain the output of the (i - 1)-th layer of the basic artificial neural network under the target intelligent node, input it into the i-th layer of the basic artificial neural network and the i-th incremental residual layer, and after fusion of the output, input it into the (i + 1)-th layer of the basic artificial neural network.

4. The task processing method according to claim 3, wherein: When the basic artificial neural network under the target intelligent node contains multiple domain labels; Obtain the output of the (i - 1)-th layer of the basic artificial neural network under the target intelligent node, and after processing by the network layer corresponding to other domain labels, fuse it into the output of the target intelligent node processing the domain task based on the incremental residual module.

5. The task processing method according to claim 3, wherein: Any incremental residual layer in the incremental residual module includes an upper backbone network label, a lower backbone network label, and a backbone network; The i-th incremental residual layer is connected to the (i + 1)-th layer of the basic artificial neural network under the target intelligent node through the upper backbone network label, and is connected to the (i - 1)-th layer of the basic artificial neural network under the target intelligent node through the lower backbone network label, and the incremental residual module is spliced with the basic artificial neural network under the target intelligent node.

6. The task processing method according to claim 1, characterized in that, Before establishing the incremental residual layer, it further includes: Determining the feature description of the task according to the domain task; Broadcasting the feature description of the task within the node community corresponding to the target domain label; When the intelligent nodes in the node community that include the incremental residual blocks corresponding to the domain task respond, obtain the incremental residual blocks according to the sharing within the node community, and replace them to generate the incremental residual module acting on the basic artificial neural network under the target intelligent node.

7. The task processing method according to claim 1, wherein After sharing the incremental residual module to other intelligent nodes in the node community corresponding to the target domain label, it further includes: For any one of the other intelligent nodes, Obtaining the new residual layer code sequence of the incremental residual module to self-check whether there is a missing incremental residual layer; If so, request to share the incremental residual module from other intelligent nodes in the node community.

8. The task processing method according to claim 1, wherein, After sharing the incremental residual module to other intelligent nodes in the node community corresponding to the target domain label, it includes Periodically checking whether there are conflicting residual layers in each intelligent node in the node community, where the conflicting residual layers are different incremental residual layers generated for the same domain task; If so, use the task evaluation function to determine the optimal incremental residual layer.

9. A task processing system for collaborative learning of multiple intelligent nodes, characterized in that, Executing the task processing method described in any one of claims 1-8 above, including: Multiple intelligent nodes to construct a target distributed intelligent network with multiple node communities, Where each intelligent node includes: A basic artificial neural network, where each network layer in the basic artificial neural network is associated with at least one domain label, and each node community corresponds to a unique domain label, so that the intelligent nodes with the same domain label form a node community; An incremental residual module, used to perform collaborative learning generation within the node community corresponding to the target domain label of the domain task, where the incremental residual module has a layer correspondence relationship with the network layer with the target domain label in the basic artificial neural network under the target intelligent node; The incremental residual module is shared to other intelligent nodes in the node community corresponding to the target domain label.

10. The task processing system according to claim 9, characterized in that: The basic artificial neural networks of each intelligent node adopt the Transformer structure, and the incremental residual blocks use the adapter module; The network layer structures corresponding to the domain labels are shared by the intelligent nodes located in the same node community.

Images