Edge Device Incremental ML Training via Weight Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing machine learning model training architecture faces privacy and security concerns due to the need to send user data to a central server, which can lead to data leaks and is inefficient, as it involves slow and periodic updates, and large file sizes that burden networks.
Innovation Solution
A system for incremental training of machine learning models on edge devices in a peer-to-peer network, where edge devices update and share model weights in real-time, enabling continuous learning without aggregating sensitive data on a central server, using a certifying node for authentication and benchmarking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If user data is sent to a central server for ML model training, then the model can be trained with aggregated data, but privacy and security concerns arise due to potential data leaks and unauthorized access
Solution Approach 1:
The patent segments the centralized training architecture into distributed edge devices that independently perform incremental training. Instead of one central server aggregating all data, multiple edge devices each train models locally using their own data, dividing the training function across many independent nodes. This segmentation eliminates the single point of failure and data concentration risk while maintaining collaborative learning through weight sharing.
Solution Approach 2:
The patent extracts the sensitive user data from the training process entirely. Rather than transmitting raw data to a server, only model weights and gradients are shared between devices. The data remains localized on each edge device, extracted from the communication stream, eliminating the data transmission and storage risks while preserving the ability to train effective models through distributed optimization.
2Productivity
If data is aggregated on a central server for periodic model updates, then training can be performed batch-wise, but the learning process becomes slow and not continuous
Solution Approach 1:
The patent implements continuous incremental training where edge devices update their models continuously as new data arrives, rather than waiting for periodic batch processing. The model weights are updated in real-time streams, and learnings are continuously shared across the network. This continuous action eliminates idle periods between batch updates and maintains constant learning progress.
Solution Approach 2:
The patent performs preliminary incremental training locally on each edge device before sharing results with the network. Each device pre-trains its model using incoming data streams, achieving preliminary learning outcomes that are then refined through weight sharing. This preliminary action at the edge eliminates the waiting time associated with centralized batch processing while maintaining collaborative improvement.
3Ease of operation
If traditional ML models are transmitted repeatedly to the server, then model updates can be deployed, but large file sizes impose significant network load
Solution Approach 1:
The patent extracts only the essential model parameters (weights and gradients) from the complete model, transmitting only these compact representations between devices. Instead of sending entire model files or raw datasets, the system transmits extracted numerical parameters that are sufficient for incremental updates. This extraction dramatically reduces transmission size while preserving the ability to perform effective model updates.
Solution Approach 2:
The patent enables each edge device to maintain and process complete model versions locally, while only exchanging compressed weight differences with the network. The full model quality resides locally at each device, while network communication carries only the minimal quality differences needed for updates. This local quality preservation eliminates repeated transmission of complete large model files.
Data Source
AI summary
A method and system for incremental training of machine learning (ML) models on edge devices, is disclosed. A base version of ML model is received by a first device of the peer to peer network of devices. The base version of the ML model is incremental trained in real-time by updating weights associated with the parameters of the ML model during a predetermined window of time. The ML model are devoid of an underlying event data used to generate the incremental version of the ML model. The updated weights associated with parameters of the incrementally trained ML model and/or the respective parameters are shared by each edge device with other edge devices. The edge devices update their respective ML models based on the received updated weights and/or parameters upon subsequent events at the edge devices and the updated version of the ML models are further incrementally trained.