General intelligence requires rethinking exploration.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cognitive Science

Background:

• Current artificial intelligence (AI) research is moving beyond simply learning from existing data.
• Large models and unified architectures like transformers highlight a new bottleneck: acquiring and utilizing relevant data effectively.
• This data acquisition challenge is framed as 'exploration', a universal learning problem in open-ended domains.

Purpose of the Study:

• To propose 'generalized exploration' as a unifying concept for exploration-driven learning across different AI paradigms.
• To highlight the essential role of exploration in all learning systems, including supervised learning, not just reinforcement learning.
• To identify key similarities and open research challenges in exploration across supervised and reinforcement learning.

Main Methods:

• Conceptual framework development to unify exploration in supervised and reinforcement learning.
• Analysis of exploration as a universal learning problem applicable to diverse AI domains.
• Identification of generalized exploration as a key objective for open-ended learning systems.

Main Results:

• Exploration is a fundamental aspect of learning essential for all AI systems, not limited to reinforcement learning.
• Generalized exploration provides a unified perspective on data acquisition and utilization across supervised and reinforcement learning.
• The proposed framework reveals commonalities and research gaps in exploration-driven AI.

Conclusions:

• Generalized exploration is crucial for enabling AI systems to continually discover and solve new problems.
• This concept is a vital step towards achieving more general artificial intelligence.
• Further research into generalized exploration can bridge current AI learning paradigms and foster more robust learning systems.