Continual-Learning Neural Control Using Pseudo-Data Rehearsal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial neural networks (ANNs) face the challenge of catastrophic forgetting, where they tend to forget previously learned information when adapting to new data, making it difficult to build upon existing models without losing previously acquired knowledge, especially in resource-constrained applications.
Innovation Solution
A control system employing two artificial neural networks, where one network generates pseudo-data samples by iteratively modifying base data samples to approach decision boundaries, allowing the other network to learn new classes while preserving existing knowledge, thereby minimizing forgetting through a method that includes partial knowledge transfer and parallel training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If an ANN is trained to learn new information, then adaptability is improved, but previously learned information is forgotten (catastrophic forgetting)
Solution Approach 1:
The system performs preliminary actions by generating pseudo-data samples that represent previously learned information before training on new data. This allows the ANN to rehearse and consolidate prior knowledge, preventing catastrophic forgetting while adapting to new information. The pseudo-data generation occurs in advance of the actual training process.
Solution Approach 2:
The system creates copies of previously learned information in the form of pseudo-data samples. These synthetic copies preserve the essential characteristics of prior knowledge and can be reused during training without requiring access to the original training data, enabling continual learning while maintaining performance on previous tasks.
2Loss of information
If historic training data is stored in a memory buffer to prevent forgetting, then information retention is improved, but device complexity and resource requirements increase
Solution Approach 1:
Instead of storing expensive, long-term memory buffers containing historic training data, the system uses inexpensive pseudo-data samples that can be generated on-demand. These synthetic samples serve as temporary, disposable representations of prior knowledge that don't require persistent storage infrastructure.
Solution Approach 2:
The system transforms the storage problem into a generation problem by changing the approach from storing actual data to generating synthetic data through parameter manipulation. The pseudo-data is created by transforming existing model parameters and learned representations rather than storing raw training examples.
3Adaptability or versatility
If the ANN architecture is simplified for resource frugal applications, then ease of deployment is improved, but the ability to handle continual learning without forgetting deteriorates
Solution Approach 1:
The system enables the ANN to serve itself by generating its own training data (pseudo-samples) from its internal representations. This self-service capability allows resource-constrained models to perform continual learning without external memory buffers or complex architectures, maintaining reliability through autonomous data generation.
Data Source
AI summary
The present disclosure relates to a control system comprising: a computation device (301) implementing at least a first artificial neural network (Net_1); one or more sensors (320) configured to capture input data samples; and one or more actuators (322); wherein the computation device (301) is configured to: generate at least one pseudo-data sample by: a) iterative sampling one of the base data samples towards one of the decision boundaries to generate one or more modified base data samples; and b) selecting one or more of the modified base data samples to form the pseudo-data sample.