System and method for training non-parametric machine learning model instances in collaborative manner

Abstract

A system and method for training non-parametric machine learning (ML) model instances in a collaborative manner is disclosed. The non-parametric ML model instances are trained on the local data set at each of the plurality of data processing nodes to obtain a plurality of non-parametric ML model instances. Each non-parametric ML model instance developed at each data processing node is shared with each of the remaining data processing nodes of the plurality of data processing nodes. Each non-parametric ML model instance of the plurality of non-parametric ML model instances is processed by a trainable parametric combiner to generate a composite model at each data processing node of the plurality of data processing nodes. Using group learning, at each data processing node of the plurality of data processing nodes, the composite model is trained on the respective local data set to obtain a trained composite model.

Description

[0001] This application is a divisional application of the invention patent application with Chinese national application number 202111245180.6, application date October 26, 2021, and invention name “System and method for collaboratively training non-parametric machine learning model instances”. Background Art

[0002] Machine learning (ML) models are trained to produce outputs based on the learning they develop on training data. During the training of an ML model, a mapping function is typically formulated to map input data to output. Based on the requirements, the performance of various ML algorithms is typically evaluated to identify the best performing ML algorithm that provides a suitable mapping function. ML algorithms that make assumptions about the form of the mapping function are used to derive parametric ML models. That is, parametric ML models learn using a predefined mapping function whose states (weights / biases) are trained using training data. Logistic regression, linear discriminant analysis, perceptron, naive Bayes, and neural networks are a few examples of parametric ML algorithms.

[0003] Non-parametric ML models do not make any assumptions about the mapping function, therefore, during the training of non-parametric ML models, the required mapping function needs to be constructed on the fly.

[0004] One way to train ML models is through collaborative learning, also known as swarm learning. Swarm learning involves collectively training ML models on local data present at different sites. That is, the ML models are trained locally on local data present at each node and provide state (weights / biases), also known as parameters, to the master node, which merges these parameters and redistributes the merged parameters to each node. Each ML model continues its learning using the merged parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] The accompanying drawings constitute part of the specification and are used to provide a further understanding of the present disclosure. Such drawings illustrate embodiments of the present disclosure for describing the principles of the present disclosure. The embodiments are shown in the accompanying drawings by way of example and not by way of limitation, and in the accompanying drawings, the same reference numerals represent the same elements. It should be noted that the "one" or "an" embodiment mentioned in the present disclosure does not necessarily refer to the same embodiment, and they mean at least one. In the drawings:

[0006] Figure 1 A network connection diagram of a blockchain system for collaboratively training a non-parametric machine learning (ML) model according to an embodiment of the present disclosure is shown;

[0007] Figure 2A block diagram showing different components of a data processing node of a blockchain system according to an embodiment of the present disclosure;

[0008] Figure 3 shows a block diagram of a data processing node illustrating development of a composite model according to an embodiment of the present disclosure;

[0009] Figure 4 shows an exemplary representation of a comparison of the output of a collaboratively developed non-parametric ML model with the output of a non-parametric ML model according to an embodiment of the present disclosure;

[0010] Figure 5 A flowchart representing a method for training a non-parametric ML model in a collaborative manner according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0011] Swarm learning can be used for parametric machine learning (ML) models. However, since non-parametric ML models do not have parameters that can be globally optimized in a collaborative manner, unlike parametric ML models, non-parametric ML models cannot be trained collaboratively.

[0012] The present disclosure provides a system and method for training non-parametric ML models in a collaborative manner while ensuring the privacy of data and ML models. As further described in the following paragraphs, in the system and method of the present disclosure, non-parametric ML models trained and developed on various data processing nodes can be shared with nodes participating in data processing in an obfuscated format. At each of the data processing nodes, each of the non-parametric ML models can be processed by a trainable parameter combiner to obtain a composite model. At each data processing node, a coordinated learning technique, i.e., group learning, can be used to train the composite model on a local data set to obtain a trained composite model. In this way, the present disclosure provides training of non-parametric ML models that use group learning to simultaneously protect the privacy of data and ML models.

[0013] By way of example, a data processing node may be responsible for training the data of a hospital using a nonparametric ML model. The data of a hospital may include health records, clinical trials, patient medical histories, administrative data, etc. Similarly, other data processing nodes may be responsible for training the data of other hospitals. It must be noted that the data of all hospitals may have the same pattern. Through training, different nonparametric ML models can be built on their respective data. However, if data from another hospital is processed, the efficiency of the nonparametric ML model will decrease. This is usually because the nonparametric ML model has not been trained for such data from another hospital. As previously described, each nonparametric ML model has its local deviation, which is learned from a local data set during training. This local deviation of the nonparametric ML model may not be suitable for processing new data and causes the nonparametric ML model to be inefficient. In order to minimize the impact of local deviations and build an efficient or excellent nonparametric ML model, group learning can be utilized. Prior to this, the nonparametric ML model needs to be processed so that it can be used for group learning.

[0014] As described above, each of the multiple data processing nodes can be responsible for processing data of corresponding hospitals located at different sites. During training, the non-parametric ML model developed at each data processing node can be shared with the remaining data processing nodes. On the one hand, the non-parametric ML model instance created at each data processing node of the multiple data processing nodes can be obfuscated before being shared with the remaining data processing nodes. Obfuscating the non-parametric ML model protects the privacy of the ML model instance. The obfuscated ML model of each data processing node in the multiple data processing nodes is then shared with adjacent data processing nodes to build a stacked collection of ML models.

[0015] At each of the plurality of data processing nodes, a corresponding stacked set of ML models may be processed by a trainable parameter combiner to generate a composite model. The trainable parameter combiner may be a parametric ML model such as a neural network that uses the predictions of the non-parametric ML model as its input and produces its predictions or inferences. That is, the output of the non-parametric ML model is treated as an input to the parametric ML model, such that a composite model is generated at each of the plurality of data processing nodes.

[0016] The composite model at each of the multiple data processing nodes may then be trained using swarm learning (or alternatively, referred to as coordinated learning techniques) at the corresponding data present locally to obtain a trained composite model.

[0017] Figure 1A network connection diagram of a blockchain system 100 for training a nonparametric machine learning (ML) model in a collaborative manner is shown. Within the blockchain system 100, data processing nodes 102, 104, 106, and 108 are interconnected. For ease of illustration and explanation, only four data processing nodes are shown, but it must be understood that a typical blockchain system will include several such data processing nodes. Each of the data processing nodes 102, 104, 106, and 108 can be responsible for training a nonparametric ML model on a respective local data set. The nonparametric ML model can be one of a decision tree, a random forest, a Bayesian model, a K-nearest neighbor (KNN), and a support vector machine (SVM).

[0018] In one embodiment, data processing node 102 may train a non-parametric ML model instance on local data set 110. Similarly, the remaining data processing nodes 104, 106, and 108 may train non-parametric ML model instances using local data sets 112, 114, and 116, respectively. Accordingly, non-parametric ML model instances F1(x), F2(x), F3(x), and F4(x) may be developed at data processing nodes 102, 104, 106, and 108, respectively.

[0019] The non-parametric ML model instance developed at the data processing node is shared with the remaining data processing nodes via the communication network 118. For example, the data processing node 102 shares its non-parametric ML model instance, i.e., Fl(x), with each of the data processing nodes 104, 106, and 108. In this way, the non-parametric ML model instances F1(x), F2(x), F3(x), and F4(x) can become available at each of the data processing nodes 102, 104, 106, and 108. In one embodiment, each data processing node can obfuscate its non-parametric ML model instance before sharing it with the remaining data processing nodes to maintain the privacy of the non-parametric ML model instance. For example, the non-parametric ML model instances F1(x), F2(x), F3(x), and F4(x) can be obfuscated using the Open Neural Network Exchange (ONNX) before they are shared between the data processing nodes 102, 104, 106, and 108. ONNX is an intermediate representation of an ML model that enables sharing of ML models from one machine learning platform to another.

[0020] In one aspect, each of the plurality of non-parametric ML model instances Fl(x), F2(x), F3(x), and F4(x) may be processed by a trainable parameter combiner to generate a composite model at each of the data processing nodes 102, 104, 106, and 108. The trainable parameter combiner may be, but is not limited to, one of logistic regression, linear discriminant analysis, perceptron, naive Bayes, and neural network. It must be carefully noted that the trainable parameter combiner is inherently trainable and parameterized.

[0021] In one embodiment, the ML model instances Fl(x), F2(x), F3(x), and F4(x) may be combined at each of the data processing nodes 102, 104, 106, and 108 using a trainable parameter combiner. The trainable parameter combiner may create an ensemble using a neural network that adjusts and evaluates the predictions of the non-parametric ML model instances. That is, the trainable parameter combiner combines the predictions of each of the non-parametric ML model instances in the non-parametric ML model instances. Because the neural network is used as a trainable parameter combiner, the weights of the network may be adjusted to further improve accuracy using group learning. On the one hand, using a trainable parameter combiner, the combination of knowledge of each non-parametric ML model instance provides a considerable advantage in reducing the impact of their respective local biases, thereby helping to build more reliable non-parametric ML model instances.

[0022] like Figure 1 As shown in , composite models 120, 122, 124, and 126 are developed using trainable parameter combiners at data processing nodes 102, 104, 106, and 108, respectively. These composite models 120, 122, 124, and 126 can be trained using group learning to obtain trained composite models at each of the data processing nodes 102, 104, 106, and 108, respectively. The parameters of such trained composite models can then be shared, aggregated, and distributed using group learning.

[0023] Figure 2 A block diagram illustrating different components of data processing node 102 is shown in accordance with an embodiment of the present disclosure. It must be understood that the structure of data processing nodes 104, 106, and 108 is similar to data processing node 102.

[0024] Data processing node 102 may include interface 202, processor 204, and memory 206. Processor 204 may be coupled to memory 206 and may be programmed to execute machine learning framework 208 and / or other instructions to perform various operations, each of which is described below.

[0025] The memory 206 may store program instructions for performing several functions by which training of non-parametric ML models may be performed by the data processing node 102. Some such program instructions stored in the memory 206 include program instructions for training non-parametric ML models, program instructions for sharing each non-parametric ML model, program instructions for generating a composite model, and program instructions for obtaining a trained composite data model using group learning.

[0026] Processor 204 may execute program instructions to train a nonparametric machine learning (ML) model through ML framework 208. The nonparametric ML model may be trained on training data 210 available locally at data processing node 102. When the nonparametric ML model is generated, the nonparametric ML model may be tested using test data 212. Nonparametric ML model instances may be created and shared with participating data processing nodes. At the same time, nonparametric ML model instances corresponding to each of the participating data processing nodes may be received at data processing node 102. Trainable parameter combiner 214 may utilize the received nonparametric ML model instances to implement respective learning of the participating data processing nodes and create a composite data model 216. The composite model 216 may be trained using group learning to obtain a trained composite model at data processing node 102. Similarly, a trained composite model will be created at each of the participating data processing nodes.

[0027] In one aspect, each of the plurality of data processing nodes 102, 104, 106, and 108 can communicate with each other for coordination. In addition, each data processing can use, for example, the blockchain API 218 to write blockchain transactions.

[0028] In another aspect, each of the plurality of data processing nodes 102, 104, 106, and 108 may register to participate in an iteration of a corresponding non-parametric ML model build. For example, the smart contract 220 may encode rules for registering a data processing node to participate in an iteration of ML model build. Thus, the blockchain system may store an immutable record of state transitions of the data processing node 102. In this manner, the blockchain system may store current and historical states of non-parametric ML models and trained composite models.

[0029] Figure 3 A block diagram of a data processing node 102 illustrating development of a composite model 120 is shown in accordance with an embodiment of the present disclosure. It must be understood that composite models 122, 124, and 126 are generated at the remaining data processing nodes 104, 106, and 108 in a similar manner.

[0030] As shown, the data processing node 102 having the non-parametric ML model instance Fl(x) developed therefrom receives the non-parametric ML model instances F2(x), F3(x) and F4(x). The receiving of the non-parametric ML model instances F2(x), F3(x) and F4(x) from the participating data processing nodes 104, 106 and 108, respectively, has been explained above. Subsequently, all the non-parametric ML model instances Fl(x), F2(x), F3(x) and F4(x) can be merged with the trainable parameter combiner 302 to produce the composite model 120. As already explained above, the trainable parameter combiner 302 is a customizable neural network. The trainable parameter combiner 302 can be selected based on a use case that enables a participant to select from predefined trainable parameter combiners or provides an application programming interface (API) to import a specific type of trainable parameter combiner.

[0031] In an exemplary aspect, the composite model 120 can be trained using group learning. As previously mentioned, collaborative training of non-parametric ML models is difficult due to the lack of parameters or predefined mapping functions. However, the composite model 120 can transform the knowledge from the shared non-parametric ML model into parameters for the trainable parameter combiner 302 to achieve prediction or inference.

[0032] When constructed, composite model 120 may be trained on the corresponding local data set, using group learning to obtain a trained composite model. Similarly, composite models 122, 124, and 126 may be constructed at data processing nodes 104, 106, and 108, and trained on their respective local data sets to obtain trained composite models. The trained composite models may combine the outputs or predictions of local training model instances conditioned by a trainable parameter combiner 302 trained on distributed data, for achieving the goal of collaborative learning of distributed data.

[0033] Figure 4 An exemplary representation showing a comparison of the output of a collaboratively developed non-parametric ML model with the output of a non-parametric data model according to an embodiment of the present disclosure. A wine dataset is processed on the blockchain system 100 to predict the quality of the wine based on factors such as fixed acidity, alcohol content, and viscosity.

[0034] Output curves 402, 404, and 406 represent the outputs of nonparametric ML model instances 1, 2, and 3, respectively, processing fixed acidity, alcohol content, and viscosity attributes. Each of the nonparametric ML model instances 1, 2, and 3 has a separate bias to estimate the data distribution of the wine data set. When the number of EPOCHs is low, the accuracy shown by the output curves 402, 404, and 406 representing the performance of the nonparametric ML model instances increases slightly, but eventually decreases or reaches a constant, i.e., below 0.85.

[0035] On the other hand, output curve 408 representing the results of the trained composite model (a non-parametric ML model developed in a collaborative manner) represents a higher accuracy than output curves 402 , 404 , and 406 representing the performance of the non-parametric ML model instances.

[0036] Now refer to Figure 5 , a method for training a non-parametric ML model in a collaborative manner is described with reference to flowchart 500. In this regard, each box may represent a module, segment, or code portion, which includes one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box may not occur in the order marked in the figure. For example, depending on the functions involved, Figure 5 Two boxes shown in succession in the flowchart may actually be executed substantially simultaneously, or the boxes may sometimes be executed in the reverse order. Any process description or box in the flowchart should be understood to represent a module, segment, or code portion, which includes one or more executable instructions for implementing specific logical functions or steps in the process. Alternative implementations are included within the scope of the example embodiments, where, depending on the functions involved, the functions may be executed in an order different from that shown or discussed, including substantially simultaneously or in the reverse order. In addition, the process description or block in the flowchart should be understood to represent decisions made by hardware structures such as state machines.

[0037] At block 502, a nonparametric ML model instance may be trained at each of a plurality of data processing nodes to obtain a plurality of nonparametric ML models. The nonparametric ML model instance may be one of a decision tree, a random forest, a Bayesian model, a K-nearest neighbor (KNN), and a support vector machine (SVM).

[0038] At block 504, each non-parametric ML model instance developed at each data processing node can be shared with each of the remaining data processing nodes of the plurality of data processing nodes. Prior to sharing, the non-parametric ML model instances can be obfuscated using the Open Neural Network Exchange (ONNX).

[0039] At block 506, each of the plurality of non-parametric ML model instances may be processed by a trainable parameter combiner to generate a composite model at each of the plurality of data processing nodes. The trainable parameter combiner may process outputs received from each of the non-parametric ML model instances. The trainable parameter combiner may be a parametric ML model such as logistic regression, linear discriminant analysis, perceptron, naive Bayes, and neural network.

[0040] At block 508 , the composite model may be trained on a corresponding local data set at each of the plurality of data processing nodes to obtain a trained composite model using swarm learning.

[0041] An embodiment of the present invention may be an article of manufacture in which a machine-readable medium (such as a microelectronic memory) (generally referred to herein as a "processor") stores thereon instructions for programming one or more data processing components to perform the above-described operations. In other embodiments, some of these operations may be performed by specific hardware components (e.g., dedicated digital filter blocks and state machines) containing hardwired logic. Those operations may alternatively be performed by any combination of programmed data processing components and fixed hardwired circuit components. In addition, while the discussion focuses on uplink media control with respect to frame aggregation, it is contemplated that control of other types of messages is also applicable.

[0042] In the above description, for the purpose of explanation, numerous specific details are set forth to provide a thorough understanding of the present system and method. It will be apparent that the system and method can be practiced without these specific details. References in the specification to "example" or similar language mean that the particular features, structures, or characteristics described in conjunction with the example are included in the scope of the description, but may not be included in other examples.

[0043] As used in this specification, the term "machine learning" refers generally to artificial intelligence techniques in which the behavior of a computer evolves based on empirical data. In some cases, the input empirical data may come from a database and produce patterns or predictions that are considered to be features of the mechanism that generated the data. In addition, the main focus of machine learning is to design algorithms that recognize complex patterns and make intelligent decisions based on the input data. Machine learning can include many methods and techniques, such as supervised learning, unsupervised learning, reinforcement learning, multivariate analysis, case-based reasoning, backpropagation, and transduction.

[0044] In some embodiments, systems and methods for training non-parametric ML models may include a physical computer-readable medium carrying instructions for processing non-parametric ML models. The computer-readable medium may be suitable for developing and training non-parametric ML models and utilizing group learning. The present disclosure maintains data privacy by processing data in a decentralized manner. Non-parametric ML models trained on multiple edge nodes may be provided to a trainable parameter combiner to output a composite model. Group learning may then process such a composite data model.

[0045] In the above description and accompanying drawings, some examples and / or implementations of systems and / or methods for training non-parametric ML models are described. As used herein, a cloud server can be used to provide services such as data processing, data communication, data storage, or any other product or activity that can run on a cloud server or is based on a cloud service. As used herein, a cloud server can be any suitable combination of physical and virtual resources pooled for computing and / or storage purposes. For example, a cloud server can include any suitable number of individual resources, servers, and server groups, including virtual instances of resources, servers, and server groups. A cloud server may include any suitable number of clouds and / or other resource networks accessible to an edge system.

[0046] The edge system may correspond to a device capable of processing / computing data received from data sources such as mobile devices, desktops, laptops, tablets, Internet of Things (IoT) devices, medical devices, industrial equipment, automobiles and other vehicles, robots, drones, etc., and may execute applications including artificial intelligence / machine learning models (i.e., data models). In some embodiments, the edge system may receive data from sensors and other interfaces to capture input data for the data model for generating prediction data, confidence scores, and performance scores.

[0047] The embodiments described above provide systems for ML model management, ML model deployment, ML model feedback collection, ML model retraining, etc. for supporting applications executable on edge systems. ML models can be selected and deployed based on characteristics shared between edge systems and other edge systems and / or cloud servers. Information received from edge systems can be used to update / retrain instances of ML models, and ML models can be tracked, recorded, and stored so that ML models can be specifically managed and customized for a single edge system, a group of edge systems, etc. ML models continuously or periodically monitor their accuracy, are updated based on the information, and are deployed to various edge systems.

[0048] The computer network may be implemented using wired and / or wireless communication technologies. The computer network may include various network components, such as switches, provider edge (PE) routers, customer edge (CE) routers, intermediate routers, bridges, computers, servers, etc. The network devices present in the computer network may implement an interior gateway protocol (IGP), including but not limited to Open Shortest Path First (OSPF), Routing Information Protocol (RIP), Intermediate System to Intermediate System (IS-IS), and Enhanced Interior Gateway Routing Protocol (EIGRP).

[0049] An interface can be used to provide input or get output from a system. An interface can be implemented as a command line interface (CLI), a graphical user interface (GUI). In addition, an application programming interface (API) can also be used to interact remotely with edge systems and cloud servers.

[0050] The processor may include one or more general purpose processors (e.g., or (AMD) microprocessors) and / or one or more special purpose processors (e.g., digital signal processors or System on a chip (SOC), Field Programmable Gate Array (FPGA) processor), MIPS / ARM-class processor, microprocessor, digital signal processor, application specific integrated circuit, microcontroller, state machine, or any type of programmable logic array.

[0051] The memory may include, but is not limited to, non-transitory machine-readable storage devices such as hard drives, magnetic tapes, floppy disks, optical disks, compact disk read-only memories (CD-ROMs) and magneto-optical disks, semiconductor memories, or other types of media / machine-readable media suitable for storing electronic instructions, such as ROM, random access memory (RAM), programmable read-only memory (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), flash memory, magnetic or optical cards.

[0052] As used herein, the terms "or" and "and / or" should be interpreted as including or meaning any one or any combination. Thus, "A, B or C" or "A, B and / or C" means "any of the following: A; B; C; A and B; A and C; B and C; A, B and C". An exception to this definition will only occur when a combination of elements, functions, steps or actions are inherently mutually exclusive in some respect.

[0053] Any combination of the above-described features and functions may be used according to one or more embodiments. In the foregoing description, embodiments have been described with reference to many specific details that may vary from implementation to implementation. Therefore, the description and drawings are to be considered illustrative rather than restrictive. The sole and exclusive indicator of the scope of the invention, and what the applicant intends to be the scope of the invention, is the text and equivalent range of the specific form in which such claims issue, as claimed in the claims issued by this application, including any subsequent corrections.

Claims

1. A method for training a non-parametric machine learning (ML) model instance, comprising: training a non-parametric ML model instance on first corresponding local data sets at a first subset of the plurality of data processing nodes to obtain a plurality of non-parametric ML model instances; sharing the plurality of non-parametric ML model instances developed at the first subset of the plurality of data processing nodes with a second subset of the plurality of data processing nodes, wherein the second subset is different from the first subset, wherein the first subset includes more than one first processing node, the second subset includes more than one second processing node, and the second processing node is different from the first processing node, and wherein the first subset and the second subset include less than all of the plurality of data processing nodes; generating, at the second subset of the plurality of data processing nodes, a set of composite models at the second subset of the plurality of data processing nodes by: processing the non-parametric ML model instances through a trainable parameter combiner; training the set of composite models at the second subset of the plurality of data processing nodes, the training utilizing second corresponding local data sets at the second subset of the plurality of data processing nodes using a coordinated learning technique to obtain a set of trained composite models, wherein the second subset of the plurality of data processing nodes independently trains the set of composite models using swarm learning; as well as An inference process of the set of trained composite models is initiated.

2. The method according to claim 1, wherein the non-parametric ML model instance is one of a decision tree, a random forest, a Bayesian model, a K-nearest neighbor (KNN), and a support vector machine (SVM).

3. The method of claim 1, wherein the trainable parameter combiner is a parametric ML model selected from at least one of logistic regression, linear discriminant analysis, perceptron, naive Bayes, and neural network.

4. The method of claim 1 , wherein the non-parametric ML model instance is obfuscated by the first subset of the plurality of data processing nodes prior to being shared with the second subset of the plurality of data processing nodes.

5. The method of claim 4, wherein at least one non-parametric ML model instance is obfuscated using an Open Neural Network Exchange (ONNX) model. The method of claim 1 , wherein the inference process is initiated at the second subset of the plurality of data processing nodes.

7. The method according to claim 1, further comprising: A set of output curves is generated, the set of output curves representing the performance of the non-parametric ML model for the set of composite models.

8. The method of claim 1, wherein the trainable parameter combiner utilizes parameters from the plurality of non-parametric ML model instances in generating the set of composite models.

9. The method according to claim 1, wherein the method further comprises: Parameters of the set of composite models are shared, aggregated, and distributed using group learning by the second subset.

10. The method of claim 1, wherein the sharing between the first subset and the second subset comprises writing blockchain transactions using a blockchain application programming interface (API).

11. A blockchain system for training a non-parametric machine learning (ML) model instance, comprising: Multiple data processing nodes, including: a first subset of the plurality of data processing nodes, and a second subset of the plurality of data processing nodes, wherein the second subset is different from the first subset, wherein the first subset includes more than one first processing node, the second subset includes more than one second processing node, and the second processing node is different from the first processing node, and wherein the first subset and the second subset include less than all of the plurality of data processing nodes; Wherein the first subset stores programmed instructions to: training a nonparametric ML model instance on first corresponding local data sets at the first subset of the plurality of data processing nodes to obtain a plurality of nonparametric ML model instances, sharing the plurality of nonparametric ML model instances developed at the first subset of the plurality of data processing nodes with a second subset of the plurality of data processing nodes; wherein the second subset stores programmed instructions to: generating, at the second subset of the plurality of data processing nodes, a set of composite models at the second subset of the plurality of data processing nodes by: processing the non-parametric ML model instances through a trainable parameter combiner, training the set of composite models at the second subset of the plurality of data processing nodes, the training utilizing second respective local models at the second subset of the plurality of data processing nodes using a coordinated learning technique to obtain a set of trained composite models, wherein each data processing node in the second subset of the plurality of data processing nodes independently trains the set of composite models using group learning; and An inference process of the set of trained composite models is initiated.

12. The blockchain system according to claim 11, wherein the non-parametric ML model instance is at least one of a decision tree, a random forest, a Bayesian model, a K-nearest neighbor (KNN), and a support vector machine (SVM).

13. The blockchain system of claim 11, wherein the trainable parameter combiner is a parameter ML model selected from at least one of logistic regression, linear discriminant analysis, perceptron, naive Bayes, and neural network.

14. The blockchain system of claim 11, wherein the non-parametric ML model instance is obfuscated by the first subset of the plurality of data processing nodes before being shared with the second subset of the plurality of data processing nodes.

15. The blockchain system of claim 14, wherein at least one non-parametric ML model instance is obfuscated using an Open Neural Network Exchange (ONNX) model.

16. A non-transitory computer-readable storage medium storing program instructions for training a non-parametric machine learning (ML) model instance, the instructions when executed on a blockchain system performing the following steps: training a non-parametric ML model instance on first corresponding local data sets at a first subset of the plurality of data processing nodes to obtain a plurality of non-parametric ML model instances; sharing the plurality of non-parametric ML model instances developed at the first subset of the plurality of data processing nodes with a second subset of the plurality of data processing nodes, wherein the second subset is different from the first subset, wherein the first subset includes more than one first processing node, the second subset includes more than one second processing node, and the second processing node is different from the first processing node, and wherein the first subset and the second subset include less than all of the plurality of data processing nodes; generating, at the second subset of the plurality of data processing nodes, a set of composite models at the second subset of the plurality of data processing nodes by: processing the non-parametric ML model instances through a trainable parameter combiner; training the set of composite models at the second subset of the plurality of data processing nodes, the training utilizing second corresponding local data sets at the second subset of the plurality of data processing nodes using a coordinated learning technique to obtain a set of trained composite models, wherein the second subset of the plurality of data processing nodes independently trains the set of composite models using swarm learning; as well as An inference process of the set of trained composite models is initiated. 17 . The non-transitory computer-readable storage medium of claim 16 , wherein the non-parametric ML model instance is one of a decision tree, a random forest, a Bayesian model, a K-nearest neighbor (KNN), and a support vector machine (SVM).

18. The non-transitory computer-readable storage medium of claim 16, wherein the trainable parameter combiner is a parametric ML model selected from at least one of logistic regression, linear discriminant analysis, perceptron, naive Bayes, and neural network.

19. The non-transitory computer-readable storage medium of claim 16, wherein a non-parametric ML model instance is obfuscated by the first subset of the plurality of data processing nodes prior to being shared with the second subset of the plurality of data processing nodes.

20. The non-transitory computer-readable storage medium of claim 19, wherein at least one non-parametric ML model instance is obfuscated using an Open Neural Network Exchange (ONNX) model.

Images