WO2023031544A1 - Procédé et système de configuration de réseaux de neurones d'un ensemble de noeuds d'un réseau de communication - Google Patents
Procédé et système de configuration de réseaux de neurones d'un ensemble de noeuds d'un réseau de communication Download PDFInfo
- Publication number
- WO2023031544A1 WO2023031544A1 PCT/FR2022/051617 FR2022051617W WO2023031544A1 WO 2023031544 A1 WO2023031544 A1 WO 2023031544A1 FR 2022051617 W FR2022051617 W FR 2022051617W WO 2023031544 A1 WO2023031544 A1 WO 2023031544A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- cluster
- nodes
- node
- weights
- model
- Prior art date
Links
- 238000004891 communication Methods 0.000 title claims abstract description 103
- 238000000034 method Methods 0.000 title claims abstract description 89
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 38
- 230000002776 aggregation Effects 0.000 claims abstract description 221
- 238000004220 aggregation Methods 0.000 claims abstract description 221
- 230000004931 aggregating effect Effects 0.000 claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 6
- 238000003062 neural network model Methods 0.000 claims abstract description 4
- 238000005192 partition Methods 0.000 claims description 30
- 230000006870 function Effects 0.000 claims description 12
- 230000005540 biological transmission Effects 0.000 claims description 2
- 238000000638 solvent extraction Methods 0.000 abstract description 3
- 239000013598 vector Substances 0.000 description 37
- 238000009826 distribution Methods 0.000 description 19
- 238000012360 testing method Methods 0.000 description 10
- 230000008859 change Effects 0.000 description 7
- 230000008521 reorganization Effects 0.000 description 7
- 238000005457 optimization Methods 0.000 description 6
- 238000004590 computer program Methods 0.000 description 4
- 230000007423 decrease Effects 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000004138 cluster model Methods 0.000 description 2
- 238000005265 energy consumption Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000004377 microelectronic Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000011664 signaling Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
Definitions
- the invention relates to telecommunication networks. It relates to learning of neural networks implemented by devices connected to a communication network.
- the invention lies more precisely in the context of federated learning in which devices locally train models of neural networks of the same structure and share the learning carried out on their devices with the other devices.
- Federated learning is opposed to centralized learning where learning is done centrally, for example on the servers of a service provider.
- Federated learning can for example be favored over centralized learning when it is difficult to envisage a global centralized model adapted to all devices.
- federated learning can also be advantageous when the devices are likely to train their models by data whose distribution is likely to depend, at least to some extent, on these devices.
- federated learning is thus a form of distributed learning where multiple nodes collaboratively solve a machine learning task.
- IID independent and identically distributed
- the invention relates to a method for configuring models of neural networks of nodes of a set of nodes of a communication network, the neural networks of said nodes all having the same structure.
- the invention relates to a method for configuring weights of models of neural networks (RN) of the same structure, of nodes of a set of nodes of a communication network, said method comprising federated learning of said weights in which said nodes locally train their neural network model and share the weights of their model with other nodes of said network, the process comprising:
- said designation comprising:
- said designation is temporary, the method comprising at least one other designation for at least one other partition of said set of nodes.
- the configuration method comprises, during said federated learning:
- the configuration method comprises a partition of all the nodes into at least one cluster taking into account a communication cost between the nodes within said at least one cluster.
- the configuration method comprises a partition of all the nodes to reorganize said clusters into at least one cluster, said reorganized clusters being constituted according to a function taking into account a communication cost between the nodes at the within a reorganized cluster and a similarity of an evolution of the weights of the models of the nodes within a reorganized cluster.
- said similarity is determined by:
- the configuration method comprises
- the configuration method includes:
- the invention relates to a coordination entity capable of configuring models of neural networks of nodes of a set of nodes of a communication network, the neural networks of said nodes all having a model of the same structure,
- the invention relates to a coordination entity capable of configuring weights of models of neural networks, of the same structure, of nodes of a set of nodes of a communication network, by federated learning of said weights in which said nodes locally train their neural network models and share the weights of their model with other nodes of said network, said coordination entity comprising at least one processor capable of:
- said designation comprising:
- the coordination entity comprises:
- said coordination entity comprises:
- - a module for sending, to a node belonging to said at least one cluster, information according to which this node must play the role of aggregation node in said cluster, and identifiers of the nodes of this cluster, said node being then qualified as the cluster's aggregation node;
- the invention relates to a learning method implemented by a node of a set of nodes of a communication network.
- the invention relates to a learning method implemented by a node of a set of nodes comprising neural networks having a model of the same structure, of a communication network, said method comprising, before a federated learning weights of said models of the neural networks of the nodes of said set, in which said nodes locally train their model of neural networks and share the weights of their model with other so-called aggregation nodes of said network:
- the learning method comprises, when said node is said aggregation node;
- the learning method comprises, when said node is said aggregation node: determining whether said cluster must be restructured by taking into account a change in the weights of said cluster and/or a change in the weights of the nodes of said cluster.
- the learning method comprises, when said node is said aggregation node; if it is determined that said cluster must be restructured, restructuring said cluster by grouping at least some of the nodes of said cluster into at least one sub-cluster, said sub-clusters being constituted according to a function taking into account a communication cost between the nodes within a said sub-cluster and a similarity of an evolution of the weights of the models of the nodes within a said sub-cluster.
- said restructuring of said cluster comprises sending, to said entity of said communication network, the identifier of an isolated node of said cluster.
- the learning method comprises, when said node is not said aggregation node:
- said method is implemented by a node belonging to a first cluster, and said entity of said communication network is:
- node of said set of nodes acting as an aggregation node managing an aggregated model of a second level cluster lower than the level of said first cluster.
- the invention relates to a learning method implemented by a node of a set of nodes of a communication network, said node being capable of playing the role of aggregation node in a cluster of nodes of the set of nodes, the nodes of this set comprising a neural network, the neural networks of these nodes all having a model of the same structure.
- This process includes:
- the invention relates to a node belonging to a set of nodes of a communication network.
- the invention relates to a node belonging to a set of nodes comprising neural networks having a model of the same structure, of a communication network, said node comprising at least one processor capable of:
- a reception from an entity of said communication network, before federated learning of the weights of said models of the neural networks of the nodes of said set, in which said nodes locally train their model of neural networks and share the weights of their model with other nodes of said network, information designating a first node of said set as an aggregation node managing an aggregated model for said federated learning and, when said node is said first node, identifiers of the nodes of said one cluster whose said aggregation node manages said aggregated model.
- the node comprises:
- an initialization module configured, in the event of reception of said learning request, to initialize the weights of the aggregated model of said cluster and the weights of the models of the nodes of said cluster with the received weights, when said node is said aggregation node;
- the invention relates to a node belonging to a set of nodes of a communication network, said node being capable of playing the role of aggregation node in a cluster of nodes of said set of nodes, the nodes of this set comprising a neural network, the neural networks of said nodes all having a model of the same structure.
- This node includes:
- an initialization module configured, in the event of reception of said learning request, to initialize the weights of the aggregated model of the cluster and the weights of the models of the nodes of said cluster with the weights received;
- the invention also relates to a system comprising a coordination entity and at least one node as mentioned above.
- the invention proposes federated learning in which nodes of the network can communicate or receive weights (or parameters) or changes in the weights of the models of their neural networks.
- These nodes can be communication devices of any type. These may include terminals, connected objects (in English loT, Internet of Things) for example cell phones, laptops, domestic equipment (for example gateways), private or public equipment , in particular of an operator of a telecommunications network, for example access points, core network equipment, servers dedicated to the invention or servers implementing operator functions for the implementation of a service in the network.
- Ni nodes can be fixed or mobile. These can be virtual machines.
- the nodes each have access to a local dataset.
- the invention can be implemented, but this in a non-limiting manner, in the context of applications or services of a communication network, for which it is not possible or desirable for the devices of the network to communicate their data either among themselves or to a centralized entity.
- a node can update the weights of its model (or more simply its model), for example by performing a gradient descent on the database of its local dataset. More precisely, a gradient descent can comprise, for a node, a calculation of the gradient of a cost function by using a certain number E of times the set of local data divided into subsets (in English batch). Hyperpara meters can be considered to parameterize a gradient descent, in particular:
- B the size of the data subset (batch), for example drawn randomly.
- the invention can be implemented with all types of data sets, for example when the data of the local data sets are not “independent and identically distributed” (IID) data, but non-IID data.
- IID independent and identically distributed
- the nodes are grouped (partitioned) into clusters (or groups of nodes), these being able to vary dynamically to help for example the convergence of the models shared by the nodes of the same cluster.
- the partition of nodes into clusters may vary, the structure of a cluster (namely in particular the set of nodes that compose it) is likely to vary over time.
- a coordination entity is configured to partition or repartition all the nodes into clusters, and to designate an aggregation node in at least some of these clusters.
- At least certain nodes of the set of nodes are capable of playing this role of aggregation node.
- the coordination entity when the coordination entity has defined a new partition of clustered nodes and designated the nodes which must play the role of aggregation node within their clusters, the coordination sends these nodes information so that they play this role of aggregation node within their cluster. It also tells them the identifiers of the nodes in the cluster.
- each node of a cluster comprises its own model, but also that each cluster comprises its own model.
- the aggregation node of a cluster manages the aggregate model of at least that cluster.
- each cluster comprises an aggregation node which manages the aggregated model of this cluster.
- the aggregated model of a cluster is obtained by aggregating the weights of the models of the nodes of the cluster trained with sets of data local to these nodes.
- the nodes of a cluster which train their models with their local datasets and which contribute to the construction of the aggregated model of the cluster can for example be qualified as worker nodes.
- a node may be able to act as an aggregation node, to act as a worker node, or to play both roles.
- the role of a node can vary over the partitions, for example be redefined with each new partition.
- the learning method is implemented by a node which, in addition to being able to play the role of aggregation node, is also capable of playing the role of working node.
- an entity of the communication network can specifically inform the node that it should act as a worker node.
- a node implicitly understands that it must play the role of working node when it receives, from an entity of the communication network, the identifier of an aggregation node of a cluster to which it belongs.
- a node When a node acts as a worker node, it receives, from the aggregation node of its cluster, weights of a model having the structure of the models of all the nodes of the set to initialize the weights of its own model and it transmits to this aggregation node the weights of its model trained with a set of data local to this node.
- the aggregation node of a cluster relays communication between nodes within the cluster.
- the cost of communication between two nodes is used as a criterion (unique or not) to determine the clusters of a partition of nodes
- the cost of communication within a cluster can be the sum of the costs communication between the cluster aggregation node and each of the cluster nodes.
- the aggregation node of a cluster is chosen close to the nodes of the cluster.
- the aggregation node of a cluster is one of the nodes of the aforementioned set of nodes. In which case, it manages not only the cluster model but also its own model as described previously.
- the aggregation node of a cluster relays the communication between the coordination entity and the nodes of its cluster.
- the aggregation node of a cluster has the possibility of reorganizing its cluster, in particular of creating sub-clusters within its cluster or of excluding nodes from its cluster .
- the model of a cluster of level n can be obtained by aggregating the models of the clusters of level n+1.
- the aggregation node of a level n cluster can for example relay the communications with the aggregation nodes of the level n-1 and/or level n+1 clusters.
- the coordination entity can be considered to be an aggregation node of the lowest level, by convention of level 0 for example.
- the network entity which sends, to a node of a level n cluster, the information according to which this node must play the said role of aggregation node in this cluster, the identifiers of the nodes of this cluster and the weights of a global model for all the nodes is: - a coordination entity as mentioned above; Or
- the network entity which sends to a node the information according to which it must play the role of working node in a level n cluster and the identifier of a aggregation node of this cluster is:
- the aggregated model of each cluster is sent to the lower level cluster, for example conditionally, such as after a constant number of iterations.
- the aggregated models can thus go back to the coordination entity which can aggregate these models in an updated version of the global model.
- This global model can then be sent back down to all the nodes for a new implementation of the method either directly or via the aggregation nodes.
- the partition of the nodes into clusters can take into account a communication cost between the nodes of at least one cluster or to take into account at least one service implemented by at least one least one of the nodes. But other criteria can be used.
- the clusters of the partition of the nodes are determined to minimize a communication cost between the nodes of this cluster. But other criteria can be used.
- the clusters of the partition can for example be determined to favor the grouping of the nodes which implement the same service in the communication network. They can also be created randomly.
- the weights of the model of a cluster can be obtained by aggregating the weights of the models of the nodes which make up this cluster.
- the nodes communicate the weights (or alternatively the gradients) of their models, resulting from local calculations from their local datasets.
- the data remains local and is not shared or transferred, which guarantees the confidentiality of the data, while achieving the learning objective.
- the invention is in this sense very different from the federated multitask optimization method described in the document "V. Smith, C. K. Chiang, M. Sanjabi, and A. Talwalkar, "Federated multi-task learning, "Advances in Neural Information Processing Systems, vol.2017-Decem, no. Nips, pp. 4425 ⁇ 1435, 2017” which does not offer to group nodes into clusters.
- Different aggregation methods can be used to update the aggregate model of a level n cluster from the aggregate models of the n+1 higher level clusters or from the models of the nodes that make up this level n cluster.
- the aggregation method used to update :
- the weights of the aggregated models of the reorganized clusters use a weighted average or a median.
- the method comprises a loop implemented within each cluster.
- the aggregated model of the cluster is communicated to each of the nodes of the cluster, each of the nodes of the cluster updates its model by performing, for example, a gradient descent with its local data and returns either its new model or the evolution or the update of its model (i.e. the difference in weights between the current iteration and the previous iteration) so that it is aggregated at the level of the aggregated model of the cluster and returned to the nodes of the cluster at the next iteration.
- This loop can comprise a constant number of iterations or not. It can for example stop when a stop condition is verified.
- the coordination entity determines how the weights of the global model evolve, for example to what extent this global model continues to converge, and decides on the advisability of redefining the clusters.
- this determination may comprise obtaining a representation of the global model in the form of a vector whose coordinates are made up of the evolutions of the weights of this model and the decision to redefine or not the clusters can take into account the norm of this vector, for example via a comparison of the norm of this vector with a constant value.
- the reorganization of the clusters is a reorganization of all the nodes into a new partition of clusters of nodes.
- new aggregation nodes can be defined for at least some of the clusters. These may be, for example, nodes of these reorganized clusters.
- the reorganized clusters are constituted according to a function taking into account:
- this similarity is determined by:
- These requests can be made to the nodes directly by the coordination entity. Alternatively, they can be performed or relayed by the aggregation nodes.
- the evolutions of the weights of the models are represented in the form of vectors and the similarity of the evolutions of the weights of the models of the different nodes is for example determined by a so-called cosine similarity method.
- the weights of the updated global model are sent back to each of the nodes, either directly or via the aggregation nodes of the clusters thus reorganized.
- the nodes can thus update their model with the global model.
- Aggregated models of reorganized clusters can also be updated with the global model.
- these new clusters are then made up of nodes selected according to a proximity criterion (communication cost for example) and whose models are likely to evolve in the same way.
- this phase may comprise a loop implemented within each reorganized cluster, identical or similar for example to that of the initialization phase.
- the aggregated model of the reorganized cluster is communicated to each of the nodes of this cluster, each of the nodes updates its model by performing a gradient descent with its local data set and returns either its new model or the evolution of its model so that it is aggregated at the level of the aggregated model of the reorganized cluster and returned to the nodes of this cluster at the next iteration.
- This loop can include a constant or variable number of iterations. It can for example stop when a stop condition is verified.
- the learning method comprises a step for determining whether at least one reorganized cluster must be restructured.
- a reorganized cluster it is determined whether a reorganized cluster must be restructured according to a convergence criterion which takes into account a change in the weights of said reorganized cluster and/or a change in the weights of the nodes of the reorganized cluster.
- a convergence criterion which takes into account a change in the weights of said reorganized cluster and/or a change in the weights of the nodes of the reorganized cluster.
- it may be a double convergence criterion taking into account a change in the weights of said reorganized cluster and a change in the weights of the nodes of the reorganized cluster.
- a reorganized cluster is determined to be restructured if the following conditions are met: (1) the evolution of the weights of said reorganized cluster is below a threshold; And
- the global model is represented in the form of a vector whose coordinates are constituted by the evolutions of the weights of this model and the norm of this vector with a numerical value, used for example as a threshold value.
- This value can be a constant or a value that depends for example on the level of the cluster or on the number of iterations already performed.
- a similarity is determined between the evolution of each of the nodes of the cluster and the evolution that the cluster would have if it were deprived of this node . For example, for a given node:
- the restructuring of a cluster comprises the grouping of at least some of the nodes of this cluster into at least one sub-cluster, these sub-clusters being constituted according to a function taking into account a cost of communication between the nodes within a said sub-cluster and a similarity of an evolution of the weights of the models of the nodes within a said sub-cluster (to minimize this function for example).
- This step is similar to the reorganization step described previously (initialization phase) except that it only applies to the nodes of the cluster to be restructured and not to all the nodes.
- isolated node of a cluster to be restructured is not assigned to a sub-cluster, this node can be assigned to another cluster.
- the aggregation node of a level n cluster when the aggregation node of a level n cluster detects an isolated node, it sends the identifier of this isolated node to an entity of the communication network so that this node is reassigned. in another cluster.
- This entity can for example be a coordination entity as mentioned below or a node which plays an aggregation role in a level n-1 cluster.
- the reassignment of an isolated node to another cluster is carried out by the coordination entity mentioned above. Therefore, in one embodiment, the configuration method includes:
- the methods mentioned above are implemented by a computer program. Consequently, the invention also relates to a computer program on a recording medium, this program being capable of being implemented by a coordination entity or more generally in a computer.
- This program includes instructions suitable for implementing a configuration method or a learning method as described above.
- These programs may use any programming language, and be in the form of source code, object code, or intermediate code between source code and object code, such as in partially compiled form, or in any other desirable form.
- the invention also relates to an information medium or a recording medium readable by a computer, and comprising instructions of a computer program as mentioned above.
- the information or recording medium can be any entity or device capable of storing the programs.
- the media may comprise a storage medium, such as a ROM, for example a CD ROM or a microelectronic circuit ROM, or even a magnetic recording medium, for example a floppy disk or a disk. hard, or flash memory.
- the information or recording medium may be a transmissible medium such as an electrical or optical signal, which may be routed via an electrical or optical cable, by radio link, by wireless optical link or by other ways.
- a program according to the invention can in particular be downloaded from an Internet-type network.
- the information or recording medium may be an integrated circuit in which a program is incorporated, the circuit being adapted to execute or to be used in the execution of one of the methods in accordance with the invention.
- FIG. 1 The figure [Fig 1] represents, in a communication network, a set of nodes that can be used in at least one embodiment of the invention
- FIG. 2 The figure [Fig 2] schematically represents a node that can be used in at least one embodiment of the invention
- FIG. 3 The figure [Fig 3] represents clusters of nodes
- FIG. 4 The figure [Fig 4] represents clusters of nodes formed on communication cost criteria
- FIG. 5 The figure [Fig 5] represents clusters of nodes formed to group together nodes whose local data have homogeneous distributions;
- FIG. 6 The figure [Fig 6] represents a vectorial representation of the evolution of the model of a node
- FIG. 7 The figure [Fig 7] represents a cluster of all the nodes of the figure [Fig. 1];
- FIG. 8 The figure [Fig 8] illustrates the use of the vector representation of the figure [Fig. 6] on the cluster of the figure [Fig 7;
- FIG. 9 The figure [Fig 9] represents a cluster initialization phase that can be implemented in at least one embodiment of the invention.
- FIG. 10 The figure [Fig 10] represents an optimization phase that can be implemented in at least one embodiment of the invention.
- FIG. 11 The figure [Fig 11] represents the hardware architecture of a coordination entity according to at least one embodiment;
- FIG. 12 The figure [Fig 12] represents the functional architecture of a coordination entity according to at least one embodiment
- FIG. 13 Figure [Fig 13] shows the hardware architecture of a node according to at least one embodiment
- FIG. 14 Figure [Fig 14] shows the functional architecture of a node according to at least one embodiment
- FIG. 15 The figure [Fig 15] illustrates performances of the invention in an exemplary implementation.
- FIG 1 represents a set of nodes N, in a geographical environment for example in a country, said nodes N, being devices capable of communicating on a communication network.
- the nodes N i each have access to a local dataset ds i .
- the data of the local data sets ds i of these nodes are non-IID data.
- the data of the nodes represented in the form of a circle come from a first distribution of independent and identically distributed data
- the data of the nodes represented in the form of a triangle come from a third distribution of independent and identically distributed data.
- the distribution of the local data ds, of a node N is not known, this being moreover liable to vary over time as the node N, acquires or generates new data and /or that some data becomes obsolete.
- Each node N can acquire or generate the data ds, from its local data set.
- These data ds can for example be communication network signaling or control data, for example quality of service data, statistics on the communication network, performance indicators of the communication network. It may also be data representative of the use of the node N,, for example durations, places or ranges of use of the node N,, data on the profiles of the users of the node N,, data on the services accessed or offered by the node N,. It can also be data acquired by the node N, or by a sensor of the node N, for example meteorological data, measurements of temperature, consumption, usage, wear, etc. It can it may also be data entered or acquired by a user of node N, for example text data (content of messages, etc.), images, videos, voice messages, audio recordings, etc.
- the local data ds, of a node N can be sensitive data in the sense that this data must not be shared or communicated to the other nodes. For example, it may be data private to a user of the node, such as personal data.
- the nodes are located, for example thanks to their GPS coordinates, and the cost of communication between two nodes is constituted by the geographical distance between these nodes.
- the communication cost between two nodes can be a measurement of throughput, latency, bandwidth of a communication between these nodes.
- a node N i comprises a neural network RN which can be trained by a learning method from the local data set ds i of this node.
- the structures (number and topology of the layers) of the models of the neural networks RN of the different nodes N are identical. But the weights (or parameters) of the models of these networks are potentially different, since these networks are trained from different local dsi datasets.
- the training of a neural network of a node N, to obtain a more efficient model can include a few iterations (or round, in English round) of gradient descent. More precisely, once the weights of the network have been initialized, during an iteration, a node N, can perform a gradient descent during E epochs (in English epoch) with different data (i.e. say, that it computes a gradient using for example each of its local data ds, E times).
- B size of the data batch (randomly drawn for example).
- the N nodes are organized into Q groups or clusters.
- each cluster Q one of the nodes N, represented in black is an aggregation node Aj of the cluster Q.
- the nodes N, within a cluster Q communicate only via the aggregation node Aj so that the communication cost between two nodes within a cluster is the sum of the communication costs between each of these nodes and the node of Aj aggregation of this cluster.
- the number of cluster levels can be arbitrary, each aggregation node of a level n greater than or equal to 1, being configured to communicate with an aggregation node of lower level n-1 with the convention introduced above.
- level 0 and 1 levels of aggregation
- level 0 levels 0 and 1
- level 0 is constituted by a node of Ao aggregation (coordination entity within the meaning of the invention).
- the figure [Fig. 3] represents the nodes of the figure [Fig. 1] grouped into two clusters Q of level 1, the aggregation nodes Aj of these clusters being configured to communicate with the aggregation node Ao of level 0.
- the dotted lines represent the clusters and the solid lines represent the communications between the knots.
- the level 0 aggregation node Ao is able to communicate directly with each of the nodes N, but in order not to overload the figure [Fig. 3,] these direct communications are not shown.
- each aggregation node Aj is configured to constitute an aggregated model for the cluster Q from the local models of the nodes N, of this cluster.
- each aggregation node of level n, n greater than or equal to 0 is configured to constitute an aggregate model of level n from the aggregate models of the clusters of level n+1.
- the level 0 aggregation node Ao is configured to constitute, optimize and reorganize the clusters and to designate an aggregation node Aj within each of these clusters;
- a level n aggregation node is configured to be able to send the weights of a model to a node either directly or via a chain of aggregation nodes of intermediate levels between the level of the aggregation node and the level of this node;
- the aggregation node Aj of a cluster Q is configured to ask the nodes N, of its cluster to update their models (for example by performing a gradient descent from their local data sets da ),
- the nodes N are configured to send to the aggregation node Aj of their cluster Q the update A0, of their model;
- an aggregation node of level n-1, n greater than or equal to 1, is configured to update its model from updates of the models of the level n clusters;
- the aggregation node Aj of a cluster Q is configured to restructure its cluster Q, for example to group certain nodes of this cluster into sub-clusters or to exclude certain nodes from its cluster Q, for example if it determines that the evolution of the models of certain nodes of its cluster does not follow that of its aggregated model or that of other nodes of its cluster.
- the clusters resulting from the partitions and the successive restructurings are formed by pursuing a dual objective, namely: limiting (for example minimizing) the communication costs between the nodes, and;
- the figure [Fig. 4] represents a grouping of the nodes N, of the figure [Fig. 1] into four clusters which optimize the communication costs, the nodes N being grouped together on a purely geographical criterion (first criterion above).
- the figure [Fig. 5] represents a grouping of the nodes N, of the figure [Fig. 1] in three clusters the nodes of the same cluster having local data sets with homogeneous distributions, in order to take into account the second criterion (to optimize for example the second criterion).
- the clusters determined by the invention can result for example from a compromise between these two organizations.
- the nodes N i communicate their local data sets ds neither to the other nodes nor to the aggregation nodes nor to the coordination entity.
- the local data sets ds therefore cannot be used directly to distribute the nodes in the clusters.
- the origin ê n represents the weights of the model before its evolution, in other words a reference model
- the norm of this vector is representative of the importance of the evolution of the model: the greater the norm, the more the model evolves;
- the direction of this vector represents the way the model evolves: if the model converges towards an optimal model for this node, the direction of this vector is directed towards this optimal model.
- a cluster Ci comprising two nodes Ni, N2 is considered, the node N2 being the aggregation node Ai of this cluster.
- the nodes Ni, N2 have local data sets dsi, ds2 from different distributions.
- FIG. 8 represents, by way of example, the weights e 0PT1 of the model considered to be optimal from the data set dsi and the weights e 0PT2 of the model considered to be optimal from the data set ds2. These weights are unknown. For the sake of simplification, it is considered in this figure that the weights e are of dimension 2, dim0 and diml.
- the vectors in dashed lines represent, at each turn t, the evolution of the model of the node Ni. We see that the norms of these vectors tend (if the model converges) to decrease at each round t, and that these vectors are (normally) directed towards the point representing the weights e 0PT1 of the optimal model of the dsi data set.
- the vectors in dotted lines represent, at each turn t, the evolution of the model of the node N2. We see that the norms of these vectors tend (if the model converges) to decrease at each round t, and that these vectors are (normally) directed towards the point representing the weights e 0PT2 of the optimal model of the data set ds2.
- the vectors in solid lines represent the evolution of the aggregated model of the cluster Ci, obtained by aggregating the models of the nodes Ni and N 2 .
- the nodes whose evolutions of the models are represented by vectors of identical or neighboring directions are intended to be grouped together in the same cluster (assuming that this grouping is not called into question by the criterion of limiting communication costs).
- the coordination entity Ao can be considered as a level 0 aggregation node.
- the cluster initialization phase comprises steps E10 to E75.
- the level 0 aggregation node Ao (coordination entity within the meaning of the invention) performs a first partition of the nodes N, to initialize the clusters Q and determines, among the nodes N,, the aggregation node Aj of each of these clusters Q.
- the method comprises a parameter kinit which defines the number of clusters which must be created during this initial partition.
- the kinit clusters are formed by only taking into account the distances between the nodes N i , on the basis of their geographical locations.
- the constitution of the clusters can for example comprise a creation of one cluster per node N, and recursive fusions of pairs of clusters closest to each other in a single cluster.
- the constitution of the clusters can for example use the Hierarchical Agglomerative Clustering algorithm presented in the document “T. Hastie, R. Tibshirani, and J. Friedma, The Elements of Statistical Learning, second edi ed. Springer, 2008”.
- the aggregation node Aj can be chosen for example as the node N, of this cluster which minimizes the sum of the distances between this node Aj and each other nodes in this cluster.
- the coordination entity Ao sends:
- the nodes N, and the aggregation node Aj of this cluster can be configured to communicate with each other.
- the coordination entity Ao on the first occurrence of a step E20, the coordination entity Ao initializes a variable t representing the number of the current round to 0 and sends to each aggregation node Aj, the weight 0° of an initial model global to all the nodes and a request to learn the models of the nodes of this cluster with these weights.
- the learning request can be accompanied by a number 8 of updates, in other words of iterations to be carried out for this learning.
- the aggregation node Aj of each of the clusters Q initializes the weights of its model 0/ for round t with the weights of the global model.
- the aggregation node Aj of each of the Q sends the weights 0/ of its aggregated model to each of the nodes N, of its cluster.
- each of the nodes N initializes the weights of its local model 0 for turn t with the weights 0/ of the model of its cluster Q.
- the aggregation node Aj when the aggregation node Aj sends the weights 0/ of the model of its cluster to a node N, it asks it to update update its model 0 .
- the aggregation node Aj communicates the hyperparameters E, B, 77 to the nodes N,.
- the node N updates its model 0 . To do this, for example, it performs a gradient descent during E epochs on a batch of size B of its local data ds,.
- the node N sends the update A0 of its model for round t to the aggregation node of its cluster Q.
- an aggregation node Aj increments the variable t (current round number) and updates the weights of the aggregated model 0/ of its cluster Cj for round t by aggregation of the updates A0 of the weights of the models of the nodes N, of this cluster Q received at step E40.
- the variables t of 0/ and of A0 differ by one unit, since for example, the weights of the model 0/ of the cluster Q for round 1 are obtained by aggregating the updates A0° of the models of the nodes N, on turn 0.
- an aggregation node Aj checks whether the 3 rounds (or iterations) have been performed, in other words whether t is divisible by 3. If this n is not the case, the result of the test E50 is negative and the aggregation node sends (during a new iteration of step E30), the weights 0/ of the model of its cluster Q updated to each of the N nodes, of its cluster. These update their model 6- (step E35) and send the update A0 to the aggregation node Aj (step E40) so that the latter increments the value t and updates the aggregated model 0/ of its cluster Q (step E45).
- the result of the test E50 is positive and the aggregation node Aj sends the aggregated model 0f of its cluster Q to a node playing an aggregation node role in a lower level cluster, namely, in this example, to the coordination entity Ao.
- the coordination entity Ao updates the weights of the global model e 1 by aggregating the aggregated models 0/ of the clusters ⁇ .
- the coordination entity Ao can use different aggregation methods depending on the embodiments, and for example the aforementioned “Federated Average” or “Coordinate-wise median” aggregation methods.
- the coordination entity Ao when the norm of the evolution A0' of its model is greater than the convergence criterion e 0 , the coordination entity Ao considers that its model continues to converge and the result of the test E60 is positive. It then sends (new occurrence of step E20), the weights 0' of the new global model to the aggregation nodes Aj of the clusters Q asking them to repeat the process described above to update their aggregated models S times 0/.
- the coordination entity Ao when the norm of the evolution A0' of its model is lower than the convergence criterion e 0 , the coordination entity Ao considers that its model no longer evolves and the result of the E60 test is negative.
- the nodes N update their model 0 during a step E65 (for example by performing a gradient descent from their local data set da) and send the update A0 of their models to the coordination entity Ao during a step E70.
- the coordination entity Ao performs a new partition of the nodes to reorganize the clusters. It is recalled that at step E10, the kinit clusters had been constituted, in the embodiment cited by way of example, by taking into account only the distances between the nodes N i , on the basis of their geographical locations .
- step E75 reorganizes the nodes into clusters so that these respond to a compromise, namely to limit (for example minimize) the communication costs between the nodes within a cluster, and on the other hand constitute clusters of nodes whose updates of their A0 models ; (t) move in the same direction, this second criterion representing a priori the fact that these nodes have local data sets from homogeneous distributions.
- step E75 reorganizes the clusters to globally optimize the dimension d, rk calculated for each pair of nodes N,, Nk: in which :
- - cii,k is a distance between updates and A0® of the models of the nodes Ni and Nk;
- step E75 the new clusters Q are reorganized and their aggregation nodes Aj are designated.
- this step E75 completes an initialization phase.
- this optimization phase comprises steps F10 to F95.
- the coordination entity Ao saves the global model as reference model 0 n .
- the coordination entity Ao sends:
- the nodes N, and the aggregation node Aj of this cluster can be configured to communicate with each other.
- the coordination entity Ao sends to each aggregation node Aj:
- the aggregation node Aj of each of the reorganized clusters Q initializes the weights of its model 0f for round t with the weights of the global model.
- the aggregation node Aj of each of the reorganized clusters Q sends the weights 0f of its aggregated model to each of the nodes N, of its cluster and their request to update their 0- model by performing gradient descent from their local dataset ds,.
- each of the nodes N initializes the weights of its local model 0- for round t with the weights 0f of the model of its cluster Q and puts this model updated by performing a gradient descent during E epochs on a batch of size B of its local data ds,.
- the node N sends the update A0 of its model for round t to the aggregation node of its cluster Q.
- the aggregation node Aj of a reorganized cluster Q determines whether the cluster Q must be restructured. In the example described here, and as described above with reference to the figure [Fig. 8], this step returns:
- the aggregation node Aj determines s 'there is at least one model of a node N, of its cluster which continues to evolve and this differently from the models of the other nodes of the cluster.
- an aggregation node Aj compares the norm of A0 with the convergence criterion e n .
- the aggregation node Aj to determine if the model of a node N, which continues to evolve, evolves differently than the models of the other nodes of the cluster Cj, the aggregation node Aj considers the angle between:
- the aggregation node Aj considers that the model of a node of the cluster evolves differently if a t > . Or :
- the aggregation node Aj of a cluster Q calculates during a step F50, the aggregated model 0f of the cluster Q obtained from the updates A0 of the model of the nodes N, of the cluster Q received at step F40.
- This updated model is sent to all the nodes Ni of the cluster Q during a new iteration of step F30.
- the loop of steps F25 to F50 is performed as long as t is less than a value T.
- Other stopping criteria can be used.
- the aggregation node Aj sends (step F58) the aggregated model 0f of its cluster Q to the coordination entity Ao (or to the lower level aggregation node).
- an aggregation node Aj determines either that the model of cluster Q no longer evolves or that there is at least one node whose model evolves in a "wrong" direction, the result of test F45 is positive, and the aggregation node Aj undertakes, during a step F60, a restructuring of the cluster Q.
- This step is similar to step E75 already described except that it only applies to the nodes of the cluster Q and not to all the nodes. It therefore produces a set of sub-clusters of the cluster SQ and of aggregation nodes SAj of these sub-clusters, these sub-clusters SQ being formed to limit the communication costs between their nodes and to group together nodes whose updates day of their models evolve substantially in the same direction. Eventually NI nodes are not assigned to any of the sub-clusters and can be considered isolated.
- the subclusters SQ and the isolated nodes NI are not treated in the same way.
- the aggregation node Aj sends:
- the aggregation node Aj sends to each aggregation node SAj:
- an aggregation node Aj creates, during a step F70, for each sub-cluster a reference model sè nj by aggregating the models of the nodes of this sub-cluster and sends then the weights of this model at the aggregation node SAj of this sub-cluster SQ.
- This sub-cluster aggregation node can then implement, recursively, the steps described above to personalize its sub-cluster.
- the lower level aggregation node namely the coordination entity Ao in this example, updates, during a step F75, the reference model ê n by aggregation of the models
- the coordination entity Ao sends the weights of this reference model to the node NI, isolated during a step F80.
- the isolated node NI initializes the weights of its local model with the weights of this reference model ê n and updates this model by performing a gradient descent.
- the isolated node NI sends the update of its model to the coordination entity Ao during a step F90.
- the isolated node NI is assigned to the cluster Q whose evolution of the model with respect to the reference model ê n , i.e. (0® - ê n ⁇ ) , the closest to A0,
- a level n aggregation node is configured to determine (step E60) if its aggregated model is still far from convergence by comparing the norm of the evolution of its model with a convergence criterion s n which may be specific to this level of aggregation.
- a level n aggregation node sends a reference model resulting from this initialization phase to the level n+1 aggregation nodes (step F10).
- Each aggregation node is configured to determine whether its cluster needs to be reconfigured (step F45) and if so, to create sub-clusters (step F60) or request the lower level aggregation node to assign a cluster (step F95) to the nodes which would be isolated.
- the coordination entity Ao has the hardware architecture of a computer. It comprises in particular a processor 11, a read only memory 12, a random access memory 13, a non-volatile memory 14 and means of communication 15.
- These means of communication 15 can in particular allow the coordination entity Ao to communicate with nodes of the network.
- the ROM 12 of the coordination entity Ao constitutes a recording medium in accordance with the invention, readable by the processor and on which is recorded a computer program PGC in accordance with the invention, comprising instructions for the performing a weight configuration method according to the invention.
- the processor of said coordination entity (Ao) can be able to:
- the program PGC defines various functional and software modules of the coordination entity Ao, capable of implementing the steps of the weight configuration method. With reference to the figure [Fig. 12], in a particular embodiment, these modules include in particular here:
- a partitioning module MP configured to partition a set of nodes N, into at least one cluster Q;
- a communication COM module configured to send to a node Aj belonging to at least one cluster Q, called aggregation node Aj, information according to which said node Aj must play the role of an aggregation node Aj in this cluster Q of nodes and identifiers of the nodes N, of this cluster Q;
- the COM module being configured to send, to at least one aggregation node Aj, a request for learning the model weights of the nodes N, of this cluster Q with the weights of a global model to all the nodes ;
- the COM module being configured to receive, from said at least one aggregation node Aj, weights of the aggregated model of the cluster Cj resulting from this learning;
- FIG. 13 represents, in one embodiment of the invention, the hardware architecture of a node Aj capable of playing the role of aggregation node in a cluster of nodes of a set of nodes, neural networks of nodes of this set all having the same structure.
- This node Aj comprises in particular a processor 21, a read only memory 22, a random access memory 23, a non-volatile memory 24 and means of communication 25.
- These means of communication 25 can in particular allow the node N to communicate with a coordination entity Ao or with other network nodes, especially within the same cluster.
- the read only memory 22 of the node Aj constitutes a recording medium in accordance with the invention, readable by the processor and on which is recorded a learning program PGA in accordance with the invention, comprising instructions for the execution of a learning method according to the invention.
- the processor of the node may be able to receive, from an entity of said communication network, before a federated learning of the weights of said models of the neural networks of the nodes of said set, in which said nodes locally train their models of neural networks and share the weights of their models with other nodes of said network of information designating a node (Aj) of said set as an aggregation node managing an aggregated model for said federated learning and, when said node is said node , identifiers of the nodes of said cluster of a cluster whose said aggregation node manages said abbreviated model.
- the program PGA defines various functional and software modules of the node Aj capable of implementing the steps of the learning method.
- these modules include in particular here:
- COM2 configured to receive, from an entity of the communication network:
- the communication module COM2 being configured to receive, from an entity of the network, a request for learning the weights of the cluster from the weights of a model having said structure; - an initialization module MIN configured, in the event of reception of said learning request, to initialize weights 0, of an aggregated model Mj of the cluster Q and of the weights e t of the models of the nodes N, of the cluster Cj with said received weights;
- an update module configured to update the weights of the aggregated model of the cluster Q, by aggregating the weights of the models of the nodes N, of the cluster Q trained with data sets local to these nodes, the weights of the models nodes of the cluster being replaced by the updated weights of the aggregated model of the cluster Q after each update;
- the communication module COM2 being configured to send to the entity of the network weights of the aggregated model of said updated cluster.
- this node then denoted N, is also capable of playing the role of working node
- the means of communication COM2 are configured for:
- FIG. 15 illustrates performance of the invention in an example implementation. More precisely :
- test images from the MNIST data set are used, which comprises images each representing a number from 0 to 9, ie ten classes. 99% accuracy means that for every 100 known new images, 99% are classified correctly.
- Parts (e) and (f) of the figure [Fig. 15] respectively illustrate the advantage of the invention in terms of communication cost reduction in the case of a federated mean type aggregation method and in the case of a median type aggregation method.
- the communication cost being in this example constituted by the sum of the communication costs (i) of the model towards the nodes and (il) back from these nodes to the aggregation nodes or to the coordination entity, taking into account the number of bits needed to send the pattern (i.e. the weights), multiplied by the sum of the distances between two knots, power a (path loss exponent).
- the single aggregation node is the barycenter of all nodes.
- cluster federated learning can help reduce the communication cost by avoiding the communication between each node in the network and the single aggregation node.
- the communication cost for the federated model by clusters varies, due to the reorganization of the clusters;
- the communication cost for the federated model by clusters hardly varies any more, the clusters being stabilized and is much lower than that of the centralized federated model.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
Claims
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP22773284.9A EP4396732A1 (fr) | 2021-08-30 | 2022-08-29 | Procédé et système de configuration de réseaux de neurones d'un ensemble de noeuds d'un réseau de communication |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
FR2109043A FR3126575A1 (fr) | 2021-08-30 | 2021-08-30 | Procédé et système de configuration de réseaux de neurones d’un ensemble de nœuds d’un réseau de communication |
FRFR2109043 | 2021-08-30 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023031544A1 true WO2023031544A1 (fr) | 2023-03-09 |
Family
ID=79170878
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/FR2022/051617 WO2023031544A1 (fr) | 2021-08-30 | 2022-08-29 | Procédé et système de configuration de réseaux de neurones d'un ensemble de noeuds d'un réseau de communication |
Country Status (3)
Country | Link |
---|---|
EP (1) | EP4396732A1 (fr) |
FR (1) | FR3126575A1 (fr) |
WO (1) | WO2023031544A1 (fr) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116502709A (zh) * | 2023-06-26 | 2023-07-28 | 浙江大学滨江研究院 | 一种异质性联邦学习方法和装置 |
CN117829269A (zh) * | 2023-12-15 | 2024-04-05 | 北京天融信网络安全技术有限公司 | 联邦学习方法、装置、计算设备及机器可读存储介质 |
CN117829274A (zh) * | 2024-02-29 | 2024-04-05 | 浪潮电子信息产业股份有限公司 | 模型融合方法、装置、设备、联邦学习系统及存储介质 |
-
2021
- 2021-08-30 FR FR2109043A patent/FR3126575A1/fr not_active Withdrawn
-
2022
- 2022-08-29 WO PCT/FR2022/051617 patent/WO2023031544A1/fr active Application Filing
- 2022-08-29 EP EP22773284.9A patent/EP4396732A1/fr active Pending
Non-Patent Citations (8)
Title |
---|
ABAD M S H ET AL: "Hierarchical Federated Learning ACROSS Heterogeneous Cellular Networks", ICASSP 2020 - 2020 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), IEEE, 4 May 2020 (2020-05-04), pages 8866 - 8870, XP033794287, DOI: 10.1109/ICASSP40776.2020.9054634 * |
BRIGGS CHRISTOPHER ET AL: "Federated learning with hierarchical clustering of local updates to improve training on non-IID data", 2020 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN), IEEE, 19 July 2020 (2020-07-19), pages 1 - 9, XP033831791, DOI: 10.1109/IJCNN48605.2020.9207469 * |
D. YINY. CHENK. RAMCHANDRANP. BARTLETT: "Byzantine-robust distributed learning: Towards optimal statistical rates", 35TH INTERNATIONAL CONFÉRENCE ON MACHINE LEARNING, vol. 13, 2018, pages 8947 - 8956 |
H. BRENDAN MCMAHANE. MOORED. RAMAGES. HAMPSONB. AGÜERAARCAS: "Communication-efficient learning of deep networks from decentralized data", PROCEEDINGS OF THE 20TH INTERNATIONAL CONFÉRENCE ON ARTIFICIAL INTELLIGENCE AND STATISTICS, vol. 54, 2017 |
MIRALLES HUGO ET AL: "Communication-efficient Federated Learning through Clustering optimization", 17 December 2021 (2021-12-17), XP055913393, Retrieved from the Internet <URL:https://hal.inria.fr/hal-03479640/document> * |
T. HASTIER. TIBSHIRANIJ. FRIEDMA: "The Elements of Statistical Learning", 2008, SPRINGER |
V. SMITHC. K. CHIANGM. SANJABIA. TALWALKAR: "Federated multi-task learning", ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS, vol. 2017, no. Nips, 2017, pages 4425 - 4435 |
Y. LECUNC. CORTES, MNIST HANDWRITTEN DIGIT DATABASE, 2010 |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116502709A (zh) * | 2023-06-26 | 2023-07-28 | 浙江大学滨江研究院 | 一种异质性联邦学习方法和装置 |
CN117829269A (zh) * | 2023-12-15 | 2024-04-05 | 北京天融信网络安全技术有限公司 | 联邦学习方法、装置、计算设备及机器可读存储介质 |
CN117829274A (zh) * | 2024-02-29 | 2024-04-05 | 浪潮电子信息产业股份有限公司 | 模型融合方法、装置、设备、联邦学习系统及存储介质 |
CN117829274B (zh) * | 2024-02-29 | 2024-05-24 | 浪潮电子信息产业股份有限公司 | 模型融合方法、装置、设备、联邦学习系统及存储介质 |
Also Published As
Publication number | Publication date |
---|---|
FR3126575A1 (fr) | 2023-03-03 |
EP4396732A1 (fr) | 2024-07-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2023031544A1 (fr) | Procédé et système de configuration de réseaux de neurones d'un ensemble de noeuds d'un réseau de communication | |
CN111309614B (zh) | A/b测试方法、装置及电子设备 | |
EP3035647B1 (fr) | Procédé de choix d'au moins un service et dispositif associé | |
US11063881B1 (en) | Methods and apparatus for network delay and distance estimation, computing resource selection, and related techniques | |
EP3414701A1 (fr) | Procédé et dispositif d'anonymisation de données stockées dans une base de données | |
CN112580814B (zh) | 可伸缩的动态转移学习机制 | |
CN115688913A (zh) | 一种云边端协同个性化联邦学习方法、系统、设备及介质 | |
WO2010034920A1 (fr) | Determination et gestion de reseaux virtuels | |
EP4042339A1 (fr) | Développement de modèles d'apprentissage automatique | |
CN116562399A (zh) | 端边云协同的模型训练方法及装置 | |
JP2023014431A (ja) | 安全な照会処理のための方法、コンピュータプログラム、およびコンピュータシステム(グラフストアにおける安全な照会処理) | |
CN113742580A (zh) | 目标类型数据的召回方法、装置、电子设备及存储介质 | |
WO2020239704A1 (fr) | Procede de gestion de ressources de telecommunications allouees dynamiquement a une pluralite d'operateurs de telecommunications, produit programme d'ordinateur et dispositifs correspondants | |
CN110310138A (zh) | 一种确定用户关系的方法及装置 | |
WO2023275462A1 (fr) | Système et procédé de gestion de données dans un environnement multi-stockage | |
FR2998697A1 (fr) | Architecture de reseau de neurone, procede d'obtention et programmes correspondants | |
FR2960732A1 (fr) | Procede de routage pseudo-dynamique dans un cluster comprenant des liens de communication statiques et programme d'ordinateur mettant en oeuvre ce procede | |
KR102722708B1 (ko) | 3D 모델의 UV 맵에 대응하는 seamless texture 생성하기 위한 방법, 서버 및 컴퓨터 프로그램 | |
US20240012942A1 (en) | Defending against adversarial attacks in federated learning | |
US20220343219A1 (en) | Parallel cross validation in collaborative machine learning | |
FR3141585A1 (fr) | Procédé d’évaluation d’un déploiement d’une configuration candidate d’un équipement d’accès radio d’un réseau de télécommunications, dispositif, système de gestion et programme d’ordinateur correspondants | |
Moreau et al. | DPSGD strategies for cross-silo federated learning | |
FR3060795A1 (fr) | Procede de construction d’une representation tridimensionnelle d’une atmosphere, dispositif et programme correspondant | |
FR3126576A1 (fr) | Procédé de détermination d’une topologie d’interconnexion entre commutateurs d’un réseau de communication, produit programme d’ordinateur et dispositif correspondants. | |
WO2022269161A1 (fr) | Procédé d'échange sécurisé de données |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22773284 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18688006 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2022773284 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2022773284 Country of ref document: EP Effective date: 20240402 |