Active Lifelong Learning With "Watchdog"

Gan Sun, Yang Cong, Xiaowei Xu

Introduction

Lifelong Learning (LL) systems aim to mimic human intelligence by sequentially acquiring knowledge from a stream of tasks, leveraging past experiences to improve performance on future ones. However, traditional LL models often assume a passive role, processing tasks in a fixed or random order. This paper introduces Active Lifelong Learning (AcLL), a framework that actively selects the most informative tasks from a candidate pool. By prioritizing "unknown" or novel tasks through a "Watchdog" mechanism, the system accumulates knowledge more efficiently and avoids the redundancy inherent in random task arrival.

Methodology: The "Watchdog" Mechanism

The core innovation of AcLL is the Watchdog mechanism, which treats task selection as an outlier detection problem. Instead of learning every task equally, the system evaluates the novelty of incoming tasks before committing resources to full training.

The Active Lifelong Learning (AcLL) framework, illustrating the interaction between the knowledge library, the hierarchical dictionary, and the task selection process.

The mechanism relies on two primary components:

  1. Hierarchical Dictionary Learning: The model employs a two-level dictionary structure ($D = D' + \Theta A$). A high-dimensional global dictionary ($D'$) captures broad patterns, while a parameterized low-dimensional local dictionary ($\Theta A$) captures shared subspaces among specific task atoms.
  2. Sparse Reconstruction Cost: When a pool of candidate tasks arrives, the Watchdog attempts to reconstruct each task using the current hierarchical dictionary. Tasks with a high reconstruction error are flagged as "unknown" (outliers) and prioritized for learning, as they represent the most significant potential for expanding the model's knowledge base.

Optimization is handled via an online update strategy using the Alternating Direction Method of Multipliers (ADMM), ensuring the dictionary evolves as new tasks are integrated without the need to store or revisit all previous raw data.

Experimental Setup

The authors evaluated AcLL on six diverse real-world benchmark datasets spanning regression and classification tasks:

  • Regression: London School Data (exam scores) and Parkinson Data (symptom prediction for "Total" and "Motor" scores).
  • Classification: Landmine Data (geophysical detection), Yeast Data (gene function), and SmartMeter Data (energy consumption).
  • Image Recognition: The Extended Yale B dataset was used to validate the hierarchical dictionary's efficiency in capturing feature distributions compared to traditional flat dictionaries.

Key Results and Insights

AcLL consistently outperformed state-of-the-art active selection baselines, such as ELLA-Info and ELLA-Diver, across all tested datasets.

Performance curves showing AUC (higher is better) and RMSE (lower is better) as a function of the number of learned tasks. The proposed AcLL method ("Ours") consistently achieves superior results.

Key findings include:

  • Superior Accuracy: On the Landmine dataset, AcLL achieved an AUC of 78.190%, significantly higher than the best baseline (77.366%).
  • Efficiency through Novelty: Using reconstruction error as a proxy for informativeness allows the model to focus on tasks that provide the most "new" information. This leads to faster knowledge accumulation and better performance with fewer training steps.
  • Hierarchical Advantage: The hierarchical dictionary structure proved more efficient than flat alternatives. On the Yale B dataset, it achieved 95.89% accuracy with fewer atoms (494) than traditional methods like K-SVD (570 atoms).

Limitations and Future Work

Despite its advantages, the AcLL framework has limitations. The dictionary update process involves matrix inversions and eigen-decompositions ($O(d^3)$), which may pose scalability challenges for extremely high-dimensional data. Additionally, the "Watchdog" requires a pool of candidate tasks to perform selection; its benefits are minimized in strictly sequential environments where tasks appear one at a time. Future research could explore more complex knowledge library structures to further enhance the transferability of learned features across increasingly diverse task domains.

Summary updated

Abstract

Lifelong learning intends to learn new consecutive tasks depending on previously accumulated experiences, i.e., knowledge library. However, the knowledge among different new coming tasks are imbalance. Therefore, in this paper, we try to mimic an effective "human cognition" strategy by actively sorting the importance of new tasks in the process of unknown-to-known and selecting to learn the important tasks with more information preferentially. To achieve this, we consider to assess the importance of the new coming task, i.e., unknown or not, as an outlier detection issue, and design a hierarchical dictionary learning model consisting of two-level task descriptors to sparse reconstruct each task with the l0 norm constraint. The new coming tasks are sorted depending on the sparse reconstruction score in descending order, and the task with high reconstruction score will be permitted to pass, where this mechanism is called as "watchdog." Next, the knowledge library of the lifelong learning framework encode the selected task by transferring previous knowledge, and then can also update itself with knowledge from both previously learned task and current task automatically. For model optimization, the alternating direction method is employed to solve our model and converges to a fixed point. Extensive experiments on both benchmark datasets and our own dataset demonstrate the effectiveness of our proposed model especially in task selection and dictionary learning.
BibTeX
@inproceedings{sun2018active,
  author = {Gan Sun and Yang Cong and Xiaowei Xu},
  title = {{Active Lifelong Learning With "Watchdog"}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11705},
}
Figure