Distributed ML Model Integration via Schema Synchronization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity and time-consuming nature of developing, deploying, and improving machine learning (ML) models, particularly neural networks, due to their difficulty in adaptation and prediction, along with the challenge of integrating multiple trained ML models trained on different data sets, leading to inconsistent predictions and high memory requirements.
Innovation Solution
A method for integrating distributed ML models by generating a non-redundant token list and IO tuples from input and output data, synchronizing models to share the same knowledge base, and transmitting this compressed form over networks, allowing for easy extension and integration of ML models without requiring large datasets to be transferred.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple ML models are trained on different training data to solve the same prediction problem, then the models can be distributed across different systems and maintain data security, but the models return different predictions for identical inputs and cannot be unified into a central prediction instance
Solution Approach 1:
The patent segments the ML model into two distinct components: a shared schema (common vocabulary, data structures, and prediction formats) that ensures consistency, and local training data that remains distributed across different systems. This segmentation allows models to be trained locally on sensitive data while maintaining unified prediction outputs through the shared schema framework.
Solution Approach 2:
The patent introduces a schema as an intermediary layer between distributed training data and prediction outputs. This schema acts as a mediator that standardizes how input data is interpreted and how predictions are formatted, enabling consistent predictions across distributed models without requiring centralization of the actual training data.
2Reliability
If training data is pooled and transferred to a central storage medium for unified ML model training, then a single central prediction instance can be created, but large datasets must be transferred over the network which increases network traffic and potentially compromises data security
Solution Approach 1:
The patent extracts the essential structural elements (schema, vocabulary, data formats) from the training data while leaving the actual sensitive data distributed locally. This extraction allows the creation of unified prediction capabilities without requiring the physical transfer or central storage of the original training datasets, thereby maintaining data security.
Solution Approach 2:
Instead of transferring original training data, the patent creates and distributes copies of the schema definition across all systems. This schema copy contains the necessary structural information for consistent predictions but excludes the sensitive training data, enabling unified behavior without data exposure risks.
3Productivity
If ML models store knowledge as weights in neural network architecture, then the models can learn complex patterns from training data, but the knowledge becomes incomprehensible to humans and difficult to manipulate or extend
Solution Approach 1:
The patent replaces the traditional neural network weight-based knowledge storage with a schema-based system using explicit data structures, vocabularies, and rule-based representations. This substitution maintains the ability to learn complex patterns while making the knowledge transparent, editable, and extendable through conventional software development practices.
Solution Approach 2:
The patent changes the fundamental parameter representation from continuous neural network weights to discrete, structured schema elements including vocabulary terms, data field definitions, and explicit relationship rules. This parameter transformation enables both automated learning and human-directed manipulation of the model knowledge.
Data Source
Figure 1A
Figure 1B~2
Figure 3
AI summary
The invention relates to a computer-implemented method for integrating distributed machine learning models, the method comprising: - providing (1402) several prediction programs (406.1, 406.2, 406.3), each containing a machine learning (ML) model (122.1, 122.2, 122.3); - training (1404) the models of the prediction programs on different training data; - synchronizing (1406) the models such that each model contains the same token list and the same set of IO tuples after synchronization; - inputting (1408) input data into one of the prediction programs; and - computing (1410) a prediction based on the input data by the one prediction program using the synchronized model of that one prediction program.