Parameter efficient collaborative learning method and system suitable for intelligent sensing device

Abstract

The application belongs to the technical field of collaborative learning, and particularly relates to a parameter efficient collaborative learning method and system suitable for intelligent sensing devices. The application is directed to a collaborative application scene of multiple distributed terminal devices with sensing, computing and communication capabilities. Under the premise of freezing backbone parameters, a shared adapter and a personalized adapter are introduced. Personalized parameters are only updated and saved locally to maintain single-device adaptation capability. Shared parameters realize knowledge sharing through interaction and weighted aggregation of neighboring devices. An adaptive weight matrix is used to fuse shared prediction and personalized prediction, so that global collaboration and personalized adaptation are considered in the distributed sensing scene with significant data distribution difference. Meanwhile, only the lightweight trainable part is updated and exchanged, which significantly reduces the end-side computing burden and communication overhead, and maintains the efficiency and convergence stability of collaborative learning under resource-limited conditions.

Description

Technical Field

[0001] This application belongs to the field of collaborative learning technology, specifically relating to an efficient collaborative learning method and system for parameters applicable to intelligent sensing devices. Background Technology

[0002] With the increasing demand for distributed intelligent sensing and collaborative decision-making, research on collaborative learning and optimization of decentralized intelligent sensing devices has received continuous attention from academia and industry in recent years. Compared with centralized architectures that rely on central servers, decentralized intelligent sensing device collaborative systems complete information dissemination and joint learning through communication between neighboring devices and local collaboration. They have natural advantages in dealing with communication bottlenecks, single points of failure, and complex network environments, and are also more in line with the "self-organizing and adaptive" system form of the edge.

[0003] Currently, distributed terminal devices with sensing, computing, and communication capabilities are being deployed more and more widely in various real-world scenarios, such as smart cars, drones, mobile robots, roadside sensing units, industrial inspection terminals, and emergency monitoring nodes. These devices are collectively referred to as intelligent sensing devices in this paper. These scenarios generally feature "multi-node parallel operation, availability of local information, and the need for collaborative improvement of overall capabilities." This necessitates that intelligent sensing device collaborative systems not only possess independent sensing and learning capabilities but also achieve stable and effective collaboration through limited communication to enhance the overall system performance and robustness.

[0004] However, in edge or distributed deployments, collaborative learning by intelligent sensing devices faces significant resource constraints. On one hand, edge devices have limited computing power, memory / GPU memory, and energy budgets, making it difficult to handle high-frequency training and updates of large-scale models. On the other hand, communication links between devices have limited bandwidth and significant latency fluctuations; frequent transmission of large-scale parameters or gradients can lead to communication congestion, decreased training efficiency, and even instability in the training process. Existing distributed collaborative training methods, under resource constraints, often require a trade-off between training overhead and collaborative benefits, making it difficult to simultaneously meet the requirements of low cost and high performance.

[0005] At the same time, intelligent sensing device collaborative systems also face a key contradiction: "it is difficult to balance globalization and personalization." Different devices have inherent differences in their environments, observation sources, and task requirements, and the distribution of local sensing data usually exhibits significant heterogeneity. If global sharing and consistency are overemphasized, cross-device interference can easily occur, causing a decline in the local performance of some devices. If local independent training is relied upon entirely, it is difficult to obtain the overall gain and generalization improvement brought about by collaborative sharing. Summary of the Invention

[0006] To address the technical challenges of high communication and computational overhead and the difficulty in balancing global collaboration with personalized adaptation in collaborative learning of decentralized intelligent sensing devices, this application provides an efficient collaborative learning method for parameters applicable to intelligent sensing devices. The technical solution is as follows:

[0007] An efficient collaborative learning method for parameters in intelligent sensing devices includes the following steps:

[0008] S1. Initialize global training parameters, set parameters and initialize each intelligent sensing device;

[0009] S2. Each intelligent sensing device calculates the predicted value of the sample based on local sensing data and locally stored parameters;

[0010] S3. The results of the shared and personalized adapter calculations are fused using an adaptive weight matrix;

[0011] S4. Fixed shared parameters, with each local polling station updating personalized adapter parameters;

[0012] S5. After all local rounds are completed, fix the personalized parameters and update the shared adapter parameters;

[0013] S6. Based on the aggregated weight matrix, each intelligent sensing device transmits the shared parameters calculated in S5 to neighboring devices, and uses the shared parameters transmitted from neighboring devices to update and iterate the local shared parameters.

[0014] Preferably, the global parameters in step S1 include: the number of intelligent sensing devices. The number of communication rounds is The number of local training rounds is And for each smart sensing device Configure local dataset Simultaneously initialize the frozen pre-trained backbone network parameters. ,in During training, the system remains frozen and does not update; simultaneously, shared adapter parameters are initialized for each smart sensing device. With personalized adapter parameters And set the learning rate for the shared adapter. Learning rate with personalized adapter Initialize the adaptive weighting matrix. This is used to coordinate the fusion contributions of personalized and shared predictions, while initializing the weight matrix for aggregation. And satisfy:

[0015] ;

[0016] in Indicates intelligent sensing devices For neighboring equipment Aggregate weights.

[0017] Preferably, step S2 is as follows:

[0018] S21. In the When the communication wheel begins, each intelligent sensing device Set the local state of the personalized adapter to ;

[0019] S22. For local training rounds Intelligent sensing devices From local perception dataset Samples (x, y) are extracted from the data, and forward prediction calculations are performed based on the shared adapter and the personalized adapter, respectively.

[0020] ;

[0021] in Indicates intelligent sensing devices The model forward mapping, To share the prediction output, For personalized prediction output, Indicates the first The shared adapter status of the wheel, Indicates the first Wheel of Life The state of the personalized adapter trained locally.

[0022] Preferably, step S3 is as follows:

[0023] S31. Fuse the two prediction outputs obtained in step S2, and obtain the fused output. satisfy:

[0024] ;

[0025] in This is an adaptive weighting matrix used to control the relative contributions of personalized outputs and shared outputs;

[0026] S32. Calculate the local loss based on the fused output:

[0027] ;

[0028] in For intelligent sensing devices The prediction loss function, This is the local loss value corresponding to the sample (x,y).

[0029] Preferably, step S4 is as follows: S41. Intelligent sensing device Local training round At that time, based on the loss obtained in step S3 Gradient estimation for constructing personalized adapters ,get:

[0030] ;

[0031] in It is a local dataset A random subset;

[0032] S42. Use learning rate Perform iterative updates on the personalized adapter parameters:

[0033] ;

[0034] S43. When completed After a round of local training, reset and store the personalized adapter parameters:

[0035] .

[0036] Preferably, step S5 is as follows:

[0037] S51. After completing the personalized adapter local update in step S4, the smart sensing device... Based on the final personalized status Construct the gradient estimate of the shared adapter using the same method as S41. ,get ;

[0038] S52. Using a learning rate Perform this round of local updates on the shared adapter to obtain the intermediate shared parameters to be aggregated:

[0039] .

[0040] Preferably, step S6 is as follows: S61. In the first step... In the aggregation phase of round-robin communication, intelligent sensing devices Share intermediate parameters Send to its neighboring device set and from each neighboring device Receive the corresponding ;

[0041] S62. Intelligent sensing devices The received intermediate shared parameters from neighboring devices are weighted and aggregated according to the weight matrix to obtain and store the parameters of the next round of shared adapters.

[0042] ;

[0043] S63. Repeat steps S2 to S6 until the termination condition is met, and output the personalized adapter parameters for each smart sensing device. With shared adapter parameters .

[0044] Preferably, during collaborative training, individual adaptability and shared knowledge transfer capability are preserved simultaneously, enabling each intelligent sensing device to form personalized parameters that conform to its own environment based on local sensing data, and to continuously obtain transferable common information through interaction with neighboring devices.

[0045] Preferably, during training, full updates and frequent transmissions of large-scale backbone parameters are avoided. Only lightweight trainable parts are updated and exchanged, which significantly reduces the computational and storage burden on the device side and reduces communication bandwidth usage. At the same time, personalized adapters are only stored locally and do not participate in cross-device exchanges.

[0046] A parameter-efficient collaborative learning system suitable for intelligent sensing devices includes intelligent sensing devices, a shared adapter, a personalized adapter, and a processing module;

[0047] Intelligent sensing devices: Calculate predicted values ​​for samples based on locally sensed data and locally stored parameters;

[0048] A personalized adapter updates personalized parameters locally to maintain compatibility with a single device.

[0049] A shared adapter enables knowledge sharing through the interaction and weighted aggregation of shared parameters with nearby devices.

[0050] Processing module: The results of the shared and personalized adapter calculations are fused using an adaptive weight matrix.

[0051] Compared with the prior art, the beneficial effects of this application are as follows:

[0052] 1) This invention introduces a shared adapter and a personalized adapter on the basis of freezing backbone parameters: personalized parameters are updated locally to maintain single-device adaptability, and shared parameters achieve knowledge sharing through interaction with neighboring devices and weighted aggregation, thereby reducing training and communication costs while taking into account overall performance improvement and individual difference adaptation.

[0053] 2) To adapt to the actual deployment conditions where edge computing and communication resources are limited, the training process avoids full updates and frequent transmissions of large-scale backbone parameters. Only lightweight trainable parts are updated and exchanged, which significantly reduces the computing and storage burden on the edge and reduces communication bandwidth usage. At the same time, the personalized parts are only stored locally and do not participate in cross-device exchanges. Without uploading the original data, the risk of sensitive information leakage is further reduced, so that the training efficiency and convergence stability of collaborative learning can still be maintained in resource-constrained scenarios. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the stages of the present invention;

[0055] Figure 2 This is a schematic diagram illustrating the specific process of the present invention. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0057] This invention proposes an efficient collaborative learning method for parameters of intelligent sensing devices, such as... Figure 1 As shown, this method introduces a shared adapter and a personalized adapter based on frozen backbone parameters: personalized parameters are updated locally to maintain single-device adaptability, while shared parameters achieve knowledge sharing through interaction with neighboring devices and weighted aggregation, thereby reducing training and communication costs while balancing overall performance improvement and adaptation to individual differences. The intelligent sensing device can be any one or more of intelligent vehicles, drones, mobile robots, roadside sensing terminals, or other devices with sensing, computing, and communication capabilities, but this invention is not limited to specific device types.

[0058] like Figure 2 As shown, an efficient collaborative learning method for parameters suitable for intelligent sensing device systems includes the following steps: (1) Initializing global training parameters:

[0059] (1.1) Set the number of devices to The number of communication rounds is The number of local training rounds is And configure a local dataset for each smart sensing device. ;

[0060] (1.2) Simultaneously initialize the frozen pre-trained backbone network parameters. ,in Keep the system frozen and do not update it during training;

[0061] (1.3) Initialize the shared adapter parameters for each smart sensing device With personalized adapter parameters And set the shared adapter learning rate. Learning rate with personalized adapter These are used to control updates to the shared adapter and the personalized adapter, respectively.

[0062] (1.4) Initialize the adaptive weighting matrix This is used to coordinate the fusion contributions of personalized and shared predictions, while initializing the weight matrix for aggregation. ,in Indicates intelligent sensing devices For neighboring equipment The aggregate weights, and satisfying .

[0063] (2) Prediction calculation process:

[0064] (2.1) in the first When the communication wheel begins, each intelligent sensing device Set the local state of the personalized adapter to ;

[0065] (2.2) For local training rounds Intelligent sensing devices From local perception dataset Samples drawn from Forward prediction calculations are performed based on the shared adapter and the personalized adapter, respectively:

[0066] ;

[0067] in Indicates intelligent sensing devices The model forward mapping, To share the prediction output, For personalized prediction output, This represents the shared adapter state in round t. Indicates the first Wheel of Life The state of the personalized adapter trained locally.

[0068] (3) Fusion prediction calculation process: (3.1) The two prediction outputs obtained in step (2) are fused, and the fused output satisfies:

[0069] ;

[0070] in This is an adaptive weighting matrix used to control the relative contributions of personalized outputs and shared outputs.

[0071] (3.2) Calculate the local loss based on the fused output:

[0072] ;

[0073] in For intelligent sensing devices The prediction loss function (e.g., cross-entropy loss). For the sample The corresponding local loss value.

[0074] (4) Personalized parameter update process:

[0075] (4.1) Intelligent sensing devices Local training round At that time, based on the loss obtained in step (3) Gradient estimation for constructing personalized adapters ,get:

[0076] ;

[0077] in It is a local dataset A random subset.

[0078] (4.2) Using the learning rate Perform iterative updates on the personalized adapter parameters:

[0079] ;

[0080] (4.3) When completed After a round of local training, reset and store the personalized adapter parameters:

[0081] .

[0082] (5) Shared parameter update process:

[0083] (5.1) After completing the local update of the personalized adapter in step (4), the smart sensing device Based on the final personalized status Construct the gradient estimate of the shared adapter using the same method as in (4.1). ,get ;

[0084] (5.2) Using the learning rate Perform this round of local updates on the shared adapter to obtain the intermediate shared parameters to be aggregated:

[0085] .

[0086] (6) Parameter aggregation and update process:

[0087] (6.1) in the In the aggregation phase of round-robin communication, intelligent sensing devices Share intermediate parameters Send to its neighboring device set and from each neighboring device Receive the corresponding ;

[0088] (6.2) Intelligent sensing devices The received intermediate shared parameters from neighboring devices are weighted and aggregated according to the weight matrix to obtain and store the parameters of the next round of shared adapters.

[0089] ;

[0090] (6.3) Repeat steps (2) to (6) until the termination condition is met, and output the personalized adapter parameters of each smart sensing device. With shared adapter parameters .

[0091] A parameter-efficient collaborative learning system suitable for intelligent sensing devices includes intelligent sensing devices, a shared adapter, a personalized adapter, and a processing module;

[0092] Intelligent sensing devices: Calculate predicted values ​​for samples based on locally sensed data and locally stored parameters;

[0093] A personalized adapter updates personalized parameters locally to maintain compatibility with a single device.

[0094] A shared adapter enables knowledge sharing through the interaction and weighted aggregation of shared parameters with nearby devices.

[0095] Processing module: The results of the shared and personalized adapter calculations are fused using an adaptive weight matrix.

[0096] Each logical step of the method described in this application can be implemented by a computer program in a real system and ultimately compiled or interpreted into a sequence of machine instructions executable by the processor. Specifically, during the initialization phase, independent address spaces and data structure descriptors are allocated in memory for frozen backbone parameters, personalized adapter parameters, and shared adapter parameters, and the frozen backbone parameters are explicitly excluded from the set of trainable parameters and the optimizer update path.

[0097] The prediction and fusion loss calculation stages mainly trigger read-only loading instructions for backbone parameters and two types of adapter parameters, as well as matrix calculation kernel calls. During the backpropagation stage, the training framework filters parameters when constructing or scheduling the computation graph, so that gradient calculation, gradient write-back, and optimizer update only apply to the gradient buffer and parameter buffer corresponding to the adapter parameters, skipping the gradient accumulation, write-back, and optimizer state update of backbone parameters.

[0098] During the personalized parameter update and shared parameter update phases, parameter update kernels are executed for both the personalized adapter and the shared adapter, and their respective optimizer states are maintained. Personalized adapter parameters are read and written only within their local address space and do not enter the cross-node communication buffer. In the parameter aggregation update phase, send and receive buffers are established for the shared adapter, and buffer copying of shared adapter parameters, writing of neighbor data, and weighted aggregation vector matrix operations based on the weight matrix are performed. After aggregation, only the aggregation result is written back to the shared adapter parameter area; cross-node transport and write-back of backbone parameters and personalized adapter parameters are not involved.

[0099] Since this application only updates and exchanges lightweight adapter parameters during training and communication, the number of related gradient propagation and parameter update instructions is reduced, and there is no need to frequently read and write large-scale backbone parameters and their optimizer states, thereby reducing memory bandwidth usage and cache pressure. At the same time, cross-node transmission objects are restricted to shared adapter parameters, reducing communication buffer read and write and data transportation overhead, making it easier to maintain efficiency and stability in training iterations under resource-constrained conditions.

[0100] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for efficient collaborative learning of parameters suitable for intelligent sensing devices, characterized in that, Includes the following steps: S1. Initialize global training parameters, set parameters and initialize each intelligent sensing device; S2. Each intelligent sensing device calculates the predicted value of the sample based on local sensing data and locally stored parameters; S3. The results of the shared and personalized adapter calculations are fused using an adaptive weight matrix; S4. Fixed shared parameters, with each local polling station updating personalized adapter parameters; Intelligent sensing devices Local training round At that time, based on the loss obtained in step S3 Gradient estimation for constructing personalized adapters Using learning rate Perform iterative updates on the personalized adapter parameters until completion. After a round of local training, the personalized adapter parameters are reset and stored; S5. After all local rounds are completed, fix the personalized parameters and update the shared adapter parameters; After completing the local update of the personalized adapter in step S4, the smart sensing device Based on the final personalized status Gradient estimation for constructing a shared adapter Using learning rate Perform this round of local updates on the shared adapter to obtain the intermediate shared parameters to be aggregated. ; S6. Based on the aggregated weight matrix, each intelligent sensing device transmits the shared parameters calculated in S5 to neighboring devices, and uses the shared parameters transmitted from neighboring devices to update and iterate the local shared parameters. S61. In the In the aggregation phase of round-robin communication, intelligent sensing devices Share intermediate parameters Send to its neighboring device set and from each neighboring device Receive the corresponding ; S62. Intelligent sensing devices The received intermediate shared parameters from neighboring devices are weighted and aggregated according to the weight matrix to obtain and store the parameters of the next round of shared adapters. ； ； in Indicates intelligent sensing devices For neighboring equipment Aggregate weights; For learning rate, For gradient estimation of the shared adapter, Indicates the first The shared adapter status of the wheel, Indicates the first Wheel of Life The state of the personalized adapter trained locally; S63. Repeat steps S2 to S6 until the termination condition is met, and output the personalized adapter parameters for each smart sensing device. With shared adapter parameters .

2. The efficient collaborative learning method for parameters of intelligent sensing devices according to claim 1, characterized in that, Step S1 global parameters include: the number of intelligent sensing devices. The number of communication rounds is The number of local training rounds is And for each smart sensing device Configure local dataset Simultaneously initialize the frozen pre-trained backbone network parameters. ,in During training, the system remains frozen and does not update; simultaneously, shared adapter parameters are initialized for each smart sensing device. With personalized adapter parameters And set the learning rate for the shared adapter. Learning rate with personalized adapter Initialize the adaptive weighting matrix. This is used to coordinate the fusion contributions of personalized and shared predictions, while initializing the weight matrix for aggregation. And satisfy: ； in Indicates intelligent sensing devices For neighboring equipment Aggregate weights.

3. The efficient collaborative learning method for parameters of intelligent sensing devices according to claim 2, characterized in that, Step S2 is as follows: S21. In the When the communication wheel begins, each intelligent sensing device Set the local state of the personalized adapter to ; S22. For local training rounds Intelligent sensing devices From local perception dataset Samples (x, y) are extracted from the data, and forward prediction calculations are performed based on the shared adapter and the personalized adapter, respectively. ； in Indicates intelligent sensing devices The model forward mapping, To share the prediction output, For personalized prediction output, Indicates the first The shared adapter status of the wheel, Indicates the first Wheel of Life The state of the personalized adapter trained locally.

4. The efficient collaborative learning method for parameters of intelligent sensing devices according to claim 3, characterized in that, Step S3 is as follows: S31. Fuse the two prediction outputs obtained in step S2, and obtain the fused output. satisfy: ； in This is an adaptive weighting matrix used to control the relative contributions of personalized outputs and shared outputs; S32. Calculate the local loss based on the fused output: ； in For intelligent sensing devices The prediction loss function, This is the local loss value corresponding to the sample (x,y).

5. The efficient collaborative learning method for parameters of intelligent sensing devices according to claim 2, characterized in that, Step S4 is as follows: S41. Intelligent sensing devices Local training round At that time, based on the loss obtained in step S3 Gradient estimation for constructing personalized adapters ,get: ； in It is a local dataset A random subset; S42. Use learning rate Perform iterative updates on the personalized adapter parameters: ； S43. When completed After a round of local training, reset and store the personalized adapter parameters: 。 6. The efficient collaborative learning method for parameters of intelligent sensing devices according to claim 5, characterized in that, Step S5 is as follows: S51. After completing the personalized adapter local update in step S4, the smart sensing device... Based on the final personalized status Construct the gradient estimate of the shared adapter using the same method as S41. ,get ; S52. Using a learning rate Perform this round of local updates on the shared adapter to obtain the intermediate shared parameters to be aggregated. .

7. The efficient collaborative learning method for parameters of intelligent sensing devices according to claim 6, characterized in that, During collaborative training, individual adaptability and shared knowledge transfer capability are preserved simultaneously, enabling each intelligent sensing device to form personalized parameters that conform to its own environment based on local sensing data, and to continuously obtain transferable common information through interaction with neighboring devices.

8. The efficient collaborative learning method for parameters of intelligent sensing devices according to claim 7, characterized in that, During training, full updates and frequent transmissions of large-scale backbone parameters are avoided. Only lightweight trainable parts are updated and exchanged, significantly reducing the computational and storage burden on the device side and reducing communication bandwidth usage. Meanwhile, personalized adapters are stored locally and do not participate in cross-device exchanges.

9. A parameter-efficient collaborative learning system suitable for intelligent sensing devices, configured in any one of the parameter-efficient collaborative learning methods for intelligent sensing devices according to claims 1-8, characterized in that, Includes intelligent sensing devices, shared adapters, personalized adapters, and processing modules; Intelligent sensing devices: Calculate predicted values ​​for samples based on locally sensed data and locally stored parameters; A personalized adapter updates personalized parameters locally to maintain compatibility with a single device. A shared adapter enables knowledge sharing through the interaction and weighted aggregation of shared parameters with nearby devices. Processing module: The results of the shared and personalized adapter calculations are fused using an adaptive weight matrix.

Images