Model Training Apparatus for Neural Network Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network models trained sequentially on different tasks or domains often experience a reduction in performance on initial tasks due to catastrophic forgetting, especially when learning new classes or domains with different input distributions, leading to a decline in model performance on previously learned tasks.
Innovation Solution
An apparatus and method that determine difference information between models and tasks/domains, generating a third model trained to perform both initial and new tasks/domains using a regularization scheme that balances performance across tasks and domains, preventing excessive bias towards new data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a neural network model is trained sequentially on different tasks or domains, then the model can learn new tasks and domains, but the performance on initial tasks deteriorates due to catastrophic forgetting
Solution Approach 1:
The patent applies preliminary action by training an initial model on the first task or domain before sequentially introducing new tasks. This establishes a foundation that can be built upon, where each subsequent model is trained on the previous model's weights plus new task data, preserving prior knowledge while adding new capabilities
Solution Approach 2:
The patent uses difference information as an intermediary element that mediates between the initial model and subsequent models. This difference information captures the essential characteristics of each new task or domain and serves as a bridge, allowing the model to adapt to new tasks while maintaining performance on previous tasks through regularization based on this difference
2Adaptability or versatility
If the model is trained on new classes with different input distributions, then the model adapts to new domains, but the performance on original tasks decreases
Solution Approach 1:
The patent applies parameter changes by modifying the training approach based on the detected difference between tasks or domains. When significant differences are detected in input distributions, the system adjusts the regularization strength and training parameters to balance adaptation to new domains with preservation of performance on original tasks
Solution Approach 2:
The patent implements feedback by continuously evaluating model performance across all previously learned tasks and domains during sequential training. The difference information serves as feedback that guides the training process, allowing the system to detect when performance on original tasks is deteriorating and adjust accordingly through regularization
Data Source
AI summary
An apparatus comprises processing circuitry configured to receive a first model and a second model; determine difference information that is representative of a difference between the first model and the second model and/or between the first task and the second task and/or between the first domain and the second domain; and generate a third model using the first model, the second model and the difference information, wherein the generating of the third model comprises training the third model to perform both of the first task and the second task and/or to operate on both the first domain and the second domain.