Abstract
Machine learning technology has been successfully applied to computer vision, natural language processing, and other fields. However, most existing models are static; their categories and parameters are fixed upon deployment, preventing them from learning new categories incrementally. This paper proposes an incremental learning method based on sample and category prototype replay to address "catastrophic forgetting" while maintaining high accuracy and low computational complexity. Our method employs a two-stage approach: sample playback using a novel SMOTE algorithm based on global sample rationality, and category prototype playback that weights samples by their "dataset abundance." Experimental results demonstrate superior performance in both prediction accuracy and forgetting rates compared to traditional baselines.
Introduction
The deployment of machine learning models in dynamic, real-world environments necessitates the ability to learn from new data streams without losing previously acquired knowledge. Traditional fine-tuning often leads to catastrophic forgetting, where updates for new classes significantly degrade performance on original ones. While replay (or playback) mechanisms—which store a subset of old data—are effective, they often suffer from category imbalance and representation sparsity. This research introduces a hybrid framework designed to mitigate these issues through rationalized oversampling and density-aware prototype calculation.
Methodology
1. Improved Sample Playback (MD-SMOTE)
To address the imbalance in replay buffers, we developed a SMOTE (Synthetic Minority Over-sampling Technique) variant based on global sample rationality. Unlike standard SMOTE, which can introduce noise at class boundaries, our approach:
- Spatial Analysis: Calculates the average distance between minority and majority samples to identify the spatial context of each sample (e.g., dense minority regions vs. boundary areas).
- Rational Interpolation: Applies specific interpolation coefficients ($\xi$) across four distinct spatial cases. This ensures that synthesized samples reinforce the underlying distribution rather than blurring class boundaries.
2. Category Prototype Playback via Dataset Abundance
Instead of relying solely on raw data storage, we utilize "prototypes"—mathematical representations of class centers. Standard methods often use simple vector averaging, which is vulnerable to outliers. Our method improves this via:
- Parzen Window Method: Used to calculate "dataset abundance" or local density for each sample.
- Weighted Averaging: Samples in high-density areas (more representative) receive higher weights, while outliers are down-weighted. This creates a robust prototype that resists "drift" as the model's feature extractor evolves over incremental stages.
Experimental Evaluation
The proposed method was evaluated against standard baselines, including Learning Without Forgetting (LWF) and Elastic Weight Consolidation (EWC), across several UCI datasets:
- Individual Household Electric Power Consumption (IDD): Tracking shifting usage patterns.
- Gas Sensor Array Drift Dataset (GSADD): Managing sensor drift over time.
- Meteorological Stations Dataset (MSD): Predicting extreme weather conditions.
Key Results
- Accuracy: On the GSADD dataset, our method maintained an average accuracy of approximately 80.73% in early incremental stages, consistently outperforming EWC and LWF.
- Forgetting Rate: The hybrid approach demonstrated a significant reduction in forgetting. In the second incremental stage, the forgetting rate was 14 percentage points lower than competing algorithms. By the fifth stage, this gap widened to 16–18 percentage points.
- Efficiency: By combining optimized SMOTE with prototype-based replay, the system maintains low computational complexity, making it viable for real-time production environments.
Discussion and Limitations
While the hybrid framework shows strong performance, certain challenges remain:
- Scalability: As the number of classes increases over many rounds, the extreme sparsity of minority samples can challenge the quality of SMOTE synthesis.
- Outlier Sensitivity: Although the Parzen window weighting mitigates noise, extreme outliers can still influence the "abundance" calculation, slightly impacting classification precision.
Conclusion
This paper presents a robust incremental learning framework that balances recent trends (via sample playback) with long-term class distributions (via weighted prototypes). By focusing on the "rationality" of replayed data and the representative "abundance" of prototypes, the method effectively combats catastrophic forgetting with minimal computational overhead. Future work will explore dynamic category-weighted playback to further refine the balance of the empirical cache as datasets evolve.
Summary updated