Technical Summary: TypiCore for Active Class-Incremental Learning on Time Series

Problem Definition

The paper addresses Active Class-Incremental Learning (ACIL) for multivariate time series classification. This setting combines two challenging constraints:

1. Class-Incremental Learning (CIL): The model must sequentially learn new classes over time while retaining the ability to classify all previously seen classes, without access to task identifiers at inference time.
2. Limited Annotation Budget: Unlike standard CIL assumptions where data streams are fully labeled, real-world time series applications (e.g., healthcare, industrial monitoring) often have automatic data acquisition but require expensive, expert-driven manual annotation.

The core challenge is to design a query strategy that selects a small subset of unlabeled samples for annotation such that the resulting memory buffer supports both plasticity (learning new classes) and stability (retaining old knowledge) under a strict labeling budget.

Methodology

The ACIL Framework

The authors propose an Active Class-Incremental Learning (ACIL) framework integrated with rehearsal-based continual learning. The process for each task $t$ involves:

• Initialization: The model inherits parameters from the previous task.
• Active Learning Cycles: The task's unlabeled dataset $D_t$ is partitioned into $q$ slices. The model performs $N_a$ cycles where, in each cycle:
  1. A Query Strategy selects a batch of $b$ samples from the current unlabeled pool.
  2. An oracle (human annotator) provides labels for these samples.
  3. The model trains on the cumulative labeled set of the current task combined with a memory retrieval batch from the previous tasks.
• Memory Consolidation: After all cycles for a task are complete, the memory buffer is updated using the newly labeled data to prepare for the next task.

The framework evaluates two underlying rehearsal mechanisms:

• Experience Replay (ER): Uses uniform random sampling for memory retrieval and reservoir sampling for memory updates.
• ASER (Adversarial Shapley Value Experience Replay): Uses importance-driven selection based on Shapley values for both retrieval and updates to prioritize informative samples.

Baseline Query Strategies Evaluated

The paper systematically evaluates three categories of query strategies:

1. Uncertainty-Based: Selects samples where the model is most uncertain (Entropy, Margin, Least Confidence).
2. Distribution-Aware (Diversity): Core-Set, which uses a greedy k-center algorithm to maximize geometric coverage of the feature space.
3. Distribution-Aware (Typicality): TypiClust, which selects high-density, prototypical samples to ensure representativeness.

Proposed Method: TypiCore

The authors identify that no single strategy achieves a robust balance between stability and plasticity. TypiCore is introduced as a hybrid query strategy that alternates between typicality-based and diversity-based selection across active learning cycles within a single task:

1. Initialization (TypiClust): The first cycle of a new task begins with TypiClust. Since the model has not yet seen labeled examples of the new classes (cold-start), selecting prototypical, high-density samples provides a stable foundation.
2. Alternation (Core-Set): Subsequent cycles alternate with Core-Set selection. This step identifies samples that maximize the geometric coverage of the feature space, ensuring the memory buffer captures diverse regions and underrepresented areas.

This strict alternation ensures the memory buffer is simultaneously representative (capturing the core structure of classes) and diverse (covering the broader feature space), preventing the memory from collapsing into a narrow subset or becoming populated only with boundary samples.

Key Results

Experiments were conducted on the TSCIL benchmark, comprising four multivariate time series datasets: UCI-HAR, UWave, GRABMyo, and WISDM.

Performance of Baselines

• Uncertainty-Based Methods: Consistently failed to outperform random selection. In many cases (e.g., Entropy on GRABMyo and WISDM), they significantly degraded performance compared to random baselines, suggesting that boundary-focused selection is misaligned with the needs of rehearsal-based learning under low budgets.
• Distribution-Aware Methods:
  • Core-Set: Achieved high plasticity (learning accuracy) but suffered from catastrophic forgetting (high Average Forgetting), as the memory became populated with atypical, boundary-adjacent samples.
  • TypiClust: Achieved high stability and low forgetting but sometimes lacked the diversity needed for optimal plasticity, particularly when paired with the ASER mechanism.

Performance of TypiCore

• Superiority: TypiCore achieved statistically significant improvements over all baselines across multiple datasets.
• Balanced Trade-off: It consistently occupied the top-right region of the Plasticity vs. Stability trade-off plots, demonstrating that it could learn new tasks effectively without sacrificing retention of old knowledge.
• Comparison to Full Supervision: On several datasets, TypiCore matched or surpassed the performance of fully supervised continual learning methods (which have access to all labels) while using only a fraction of the available labels (e.g., 1% to 30% annotation budgets).
• Replay Mechanism Interaction: A consistent finding was that ER (simple random rehearsal) outperformed ASER (complex importance-driven rehearsal) when combined with active learning strategies, particularly under severe annotation budgets. The complexity of ASER appeared less robust when the labeled pool was small and potentially unrepresentative in early cycles.

Significance and Claims

The paper claims that TypiCore provides a novel solution to the label-scarce setting in time series continual learning. Its significance lies in:

1. Demonstrating the Limitations of Standard Strategies: It reveals that standard uncertainty-based methods are detrimental in ACIL settings and that single-objective distribution-aware methods (pure diversity or pure typicality) are insufficient.
2. Hybrid Strategy Efficacy: It proves that dynamically alternating between typicality and diversity selection creates memory buffers that are both robust against forgetting and expressive enough to capture data complexity.
3. Practical Applicability: The method enables high-performance continual learning on time series data with significantly reduced annotation costs, making it viable for real-world domains where expert labeling is a bottleneck.

The authors conclude that intelligent label selection acts as an implicit regularizer, filtering out noisy or redundant samples that would otherwise destabilize the learning process, and that simpler rehearsal mechanisms (ER) may be more compatible with active learning constraints than complex value-driven ones (ASER).