A Comparative Study on Learning Approaches towards Artificial General Intelligence

Authors

  • Aditya Bhowmic Department of Computer Science & Engineering Sharda School of Computing Science and Engineering Sharda University, Greater Noida. Author
  • Amit Sharma Department of Computer Science & Engineering Sharda School of Computing Science and Engineering Sharda University, Greater Noida. Author

DOI:

https://doi.org/10.64764/g4yn8k37

Keywords:

Artificial General Intelligence (AGI), Continual Learning, Catastrophic Forgetting, Self-Supervised Learning, Meta-Learning, Elastic Weight Consolidation (EWC), Experience Replay

Abstract

The concept of AGI (Artificial General Intelligence) denotes systems able to perform a diverse set of tasks in a range of environments, rather than a single narrow task. Among self-supervised prediction scaling, reward-based interaction, meta-learning, structured world models and continual learning, which strategies appear to be most promising in providing such generality? This work is a theoretical comparative study. It first puts the literature in order by presenting a taxonomy of the relationships of these paradigms to the requirements of general intelligence and its main constraint. Then it investigates a problem they all share in common: that is the inability to learn new skills sequentially while maintaining prior ones (catastrophic forgetting). The continual learning problem is formalized, standard three scenarios (task-, domain- and class-incremental) are described and main families of mitigations, regularisation (Elastic Weight Consolidation) and replay (experience replay) are analysed and compared with the upper bound of joint training. On the basis of the published evidence and the problem structure, it is argued that weight regularisation alone is insufficient for the hardest case (class-incremental) and that an explicit memory mechanism should be involved, according to the principle of complementary-learning-systems from neuroscience. The study concludes that a single learning approach is not sufficient for general intelligence, and advocates that hybridisation of large-scale learning with efficient memory and adaptation appears most justifiable. Essentially there are two types of AIs: general AI which can learn and operate in many different contexts (General AI), and specific AI which can only work in one specific task (specific AI).

0 0

Downloads

Published

2026-10-12

How to Cite

Aditya Bhowmic, & Amit Sharma. (2026). A Comparative Study on Learning Approaches towards Artificial General Intelligence. International Journal of Innovations in Engineering Technology and Applied Sciences, 2(2), 1-8. https://doi.org/10.64764/g4yn8k37