Meta-Learning Frameworks for Lifelong Artificial General Intelligence: Enabling Autonomous Knowledge Acquisition Across Diverse Domains

Abhishek Saxena, Sneha Saxena, Gaurav Tripathi

Abstract


Artificial General Intelligence (AGI) requires autonomous, continuous, and transferable knowledge acquisition across highly heterogeneous problem domains without suffering from catastrophic forgetting or computational stagnation. Traditional deep learning paradigms, while effective within isolated tasks, exhibit brittle transferability and rapid performance degradation when exposed to sequential non-stationary environments. Meta learning, or 'learning to learn,' has emerged as a foundational theoretical framework for constructing lifelong AGI systems. By abstracting optimization mechanics, feature representation spaces, and inference rules into meta-level priors, meta-learning enables neural agents to rapidly adapt to novel tasks using minimal target data. This comprehensive review paper synthesizes state of-the-art meta-learning methodologies tailored for lifelong autonomous learning. We systematically analyze optimization-based, model-based, and metric-based meta-learning algorithms, evaluating their structural mechanisms for mitigating catastrophic forgetting, facilitating bi-directional knowledge transfer, and maintaining dynamic memory efficiency. Furthermore, we establish a rigorous taxonomy connecting lifelong learning primitives with meta-learning architectures, highlight prevailing research gaps in memory indexing and continuous meta-optimization, and outline actionable future research directions. Our comparative analysis across multi domain benchmarks demonstrates that integrating episodic memory buffer dynamics with meta-regularization yields substantial improvements in forward knowledge transfer (up to 24.5% task performance gain) while reducing catastrophic forgetting by over 62% compared to standard continual learning baselines.

KEYWORDS: Artificial General Intelligence (AGI), Meta-Learning, Lifelong Learning, Continual Knowledge Acquisition, Catastrophic Forgetting, Meta Optimization, Epistemic Memory.


Full Text:

PDF 83-96

Refbacks

  • There are currently no refbacks.