Reservoir Management for Online Neural Learning With Balanced Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models face challenges in maintaining accurate recognition performance due to catastrophic forgetting and imbalanced label distributions in time-series data streams, particularly for minor classes, leading to gradient dominance and under-representation.
Innovation Solution
A reservoir management method that dynamically adjusts the neural network's training data by adding input data based on a sampling probability and removing candidate data to maintain a balanced label distribution, using a target label distribution and distance vectors to optimize memory allocation and training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If neural network models are trained using all available training data, then the model can learn from comprehensive data, but catastrophic forgetting occurs and recognition performance degrades over time
Solution Approach 1:
The patent extracts only the essential and representative training data from the available data stream, storing them in a reservoir. This selective extraction prevents catastrophic forgetting by maintaining a curated subset of training data that preserves recognition performance while avoiding the burden of processing all available data.
Solution Approach 2:
The patent performs preliminary selection of training data before actual model training by maintaining a reservoir of pre-selected representative samples. This preliminary action ensures that when training is needed, the model has immediate access to high-quality training data without requiring reprocessing of the entire data stream.
2Adaptability or versatility
If the reservoir stores diverse sample data, then the model can learn various patterns, but the label distribution becomes imbalanced with under-representation of minor classes
Solution Approach 1:
The patent applies local quality by managing different classes of data with different characteristics. It specifically addresses the label distribution imbalance by ensuring minor classes are adequately represented in the reservoir through targeted sampling strategies, while maintaining overall diversity for pattern recognition capability.
Solution Approach 2:
The patent changes the sampling parameters dynamically based on the current label distribution in the reservoir. When minor classes are under-represented, the sampling probability for those classes is increased, thereby adjusting the parameter distribution to achieve better balance while maintaining adaptability.
3Adaptability or versatility
If the reservoir is continuously updated with new data, then the model adapts to new patterns, but memory resources are consumed and training efficiency decreases
Solution Approach 1:
The patent applies partial action by updating only a portion of the reservoir with new data rather than completely replacing or extensively processing all data. This selective update maintains adaptability to new patterns while consuming fewer memory resources and preserving training efficiency.
Solution Approach 2:
The patent ensures continuous useful action by maintaining the reservoir as a persistent data structure that is gradually updated with new samples. This continuous update process enables online learning capability while managing memory resources efficiently through incremental rather than batch updates.
4Productivity
If sampling probability is used to select data for the reservoir, then data selection becomes efficient, but the selection process becomes complex with multiple parameters
Solution Approach 1:
The patent uses parameter changes in the sampling probability calculation to balance efficiency and complexity. By dynamically adjusting sampling probabilities based on simple criteria such as class frequency and reservoir composition, the system achieves efficient data selection without requiring overly complex selection mechanisms.
Data Source
AI summary
A reservoir management method includes: in response to receiving input data to which label information is mapped, determining whether to add the input data to a reservoir based on a sampling probability; in response to determining to add the input data to the reservoir when the reservoir is filled, selecting candidate data to be removed from among sets of sample data included in the reservoir based on a target label distribution and a current label distribution of the reservoir, and removing the selected candidate data from the reservoir; and training a neural network model using sample data of the reservoir from which the selected candidate data is removed.