New dual-stream method tackles forgetting in time series AI
Combining deep temporal embeddings and statistical features for class-incremental learning.
Pablo García-Santaclara and colleagues introduce a novel method for class-incremental continual learning tailored to multivariate time series classification. The core innovation is a dual-stream feature extraction pipeline: one stream uses a pre-trained frozen foundation model to generate deep temporal embeddings, while the other extracts traditional statistical features (like mean, variance, and autocorrelation). These are combined and fed into a rehearsal-based classifier that replays a small buffer of old data to prevent catastrophic forgetting.
Evaluated on five benchmark datasets (including UCR and UEA archives), the system shows competitive average accuracy across all configurations while maintaining low forgetting rates—even as new classes are added incrementally. This work is especially relevant for real-world applications like sensor fault detection, activity recognition, and financial forecasting, where models must adapt to new patterns without retraining from scratch. The paper is available on arXiv under the identifier 2606.03292.
- Dual-stream approach merges deep temporal embeddings (from frozen foundation model) with statistical features for richer representations.
- Tested on 5 benchmark multivariate time series datasets with competitive accuracy and low forgetting across all setups.
- Built for class-incremental scenarios where models must learn new categories without forgetting old ones.
Why It Matters
Enables AI systems to continuously adapt to new time series patterns without retraining, reducing data and compute costs.