Federated Model Aggregation to Prevent Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning systems face challenges in integrating new knowledge from devices that join after model training due to catastrophic forgetting, where neural networks abruptly forget previous information upon learning new data, and existing incremental learning methods are unsuitable or inefficient for federated setups.
Innovation Solution
An artificial intelligence aggregation system that includes a coordinator to aggregate updates from multiple sources, distributing an aggregated updated model or initial model to new sources, using a private version of replay methods to apply existing learned structures and avoid catastrophic forgetting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a neural network is retrained on new data to improve robustness, then the model can learn from new information, but the actual accuracy of the machine learning model deteriorates due to catastrophic forgetting
Solution Approach 1:
The patent creates a copy of the historical model parameters and uses this copy as a reference during incremental training. The historical model copy serves as a memory that prevents the current model from forgetting previous knowledge while learning new information, thereby maintaining accuracy while adapting to new data
Solution Approach 2:
The patent introduces an intermediary mechanism (model copying and parameter management) between the historical model and the incremental training process. This intermediary allows the system to preserve historical knowledge while enabling continuous learning, resolving the conflict between adaptability and reliability
2Adaptability or versatility
If incremental learning is used to continuously train the model on new data, then the model can adapt to new information, but catastrophic forgetting occurs where previously learned information is lost
Solution Approach 1:
The patent performs preliminary action by creating and storing copies of the historical model parameters before incremental training begins. These pre-stored model copies serve as a foundation that prevents information loss during subsequent continuous training operations
Solution Approach 2:
The system creates copies of the historical neural network model parameters and maintains these copies throughout the incremental learning process. This copying mechanism ensures that historical knowledge is preserved while the model continues to learn from new data streams
3Quantity of substance
If federated learning is used to train models across decentralized devices, then access to large amounts of data is achieved, but integrating new knowledge becomes challenging when devices join after initial training
Solution Approach 1:
The patent implements a universal model distribution mechanism where the same model copying and distribution infrastructure serves both initial federated training and subsequent integration of new devices. This multi-functional approach enables seamless onboarding of new devices without requiring separate integration processes
Solution Approach 2:
The system introduces an intermediary model management layer that handles both initial model distribution and subsequent updates to new devices. This intermediary coordinates the flow of model parameters between the central server and all devices in the federation, ensuring consistent knowledge integration
Data Source
AI summary
An artificial intelligence aggregation system (110) includes a computer (500) and a memory system. The computer (500) includes a memory (520) that stores instructions and a processor (510) that executes the instructions. The memory system aggregates (S326) a first set of updates to an initial model in a federated learning process. The computer (500) executes the instructions to: distribute (FIG. 3B), to sources of the first set of updates in a federation, a first aggregated updated model that aggregates updates to the initial model; and distribute (FIG. 3B), to a first new source, either the initial model or the first aggregated updated model.