Meta-learning Network Resolves Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models typically experience catastrophic forgetting, where they lose previously acquired knowledge when trained on new data or tasks, especially under few-shot learning paradigms.
Innovation Solution
The technique involves performing meta-learning by converting a prediction learning network and a representation learning network into trained networks based on support and query sets of training data, using both supervised and self-supervised training outputs to generate latent representations without labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a machine learning model is trained on new data or tasks using conventional machine learning workflows, then the model can learn new concepts and tasks, but the model experiences catastrophic forgetting and loses previously acquired knowledge
Solution Approach 1:
The system performs preliminary action by training an initial machine learning model on a first task before encountering new tasks. This establishes a baseline of learned knowledge that subsequent techniques aim to preserve. The meta-learning framework is pre-configured with this initial training, creating a foundation that can be built upon without complete retraining.
Solution Approach 2:
The system implements feedback through the evaluation function that assesses model performance on both new tasks and previously learned tasks. This feedback mechanism allows the system to detect forgetting and adjust training parameters accordingly. The loss function incorporates performance on prior tasks, creating a feedback loop that prevents catastrophic forgetting by continuously monitoring and maintaining previously acquired knowledge.
2Productivity
If a machine learning model is trained under a few-shot learning paradigm with rapid parameter changes to adapt to limited data, then the model can quickly adapt to new tasks with few samples, but the rapid parameter changes interfere with the ability to retain knowledge from previous tasks
Solution Approach 1:
The system performs preliminary action by pre-training the model on a substantial first task before entering the few-shot learning paradigm. This preliminary training establishes robust feature representations and knowledge structures that are more resistant to catastrophic forgetting. The meta-learning framework is prepared in advance with this foundation, enabling faster adaptation to subsequent few-shot tasks while preserving earlier knowledge.
Solution Approach 2:
The system implements feedback through the composite loss function that combines performance on new tasks with performance on previously learned tasks. This feedback mechanism ensures that rapid parameter changes during few-shot learning do not completely overwrite previous knowledge. The evaluation function continuously monitors both new and old task performance, providing feedback that balances adaptation speed with knowledge retention.
3Adaptability or versatility
If a machine learning model is retrained to perform sequences of tasks using conventional workflows, then the model can be updated with new tasks, but the model loses accuracy on previously learned tasks due to parameter overwriting
Solution Approach 1:
The system performs preliminary action by establishing a meta-learning framework before sequence learning begins. This framework includes pre-training on the first task and configuring the loss function to account for multiple tasks. The preliminary setup creates a structure that can accommodate sequential task learning while preserving accuracy on previously learned tasks through the designed feedback mechanisms.
Solution Approach 2:
The system implements feedback through the multi-task loss function that simultaneously considers performance on current and previous tasks. This feedback mechanism prevents parameter overwriting by adjusting weights and parameters in a way that maintains accuracy across the task sequence. The evaluation function provides continuous feedback on both new and old task performance, enabling the model to adapt to sequences of tasks while preserving accuracy through iterative refinement.
Data Source
AI summary
One embodiment of the present invention sets forth a technique for performing meta-learning. The technique includes performing a first set of training iterations to convert a prediction learning network into a first trained prediction learning network based on a first support set of training data and executing a representation learning network and the first trained prediction learning network to generate a first set of supervised training output and a first set of self-supervised training output based on a first query set of training data corresponding to the first support set of training data. The technique also includes performing a first training iteration to convert the representation learning network into a first trained representation learning network based on a first loss associated with the first set of supervised training output and a second loss associated with the first set of self-supervised training output.