Model capacity's role in grokking explained by competing memorization and generalization speeds

A new research paper proposes an information-theoretic framework to explain how model capacity influences the phenomenon of "grokking" in machine learning. The study suggests that grokking, where a model suddenly generalizes after initially memorizing training data, emerges from a competition between two distinct timescales: memorization speed and generalization speed. These timescales are shown to be dependent on the model's parameter count, with larger models exhibiting faster memorization. AI

IMPACT Provides a theoretical framework for understanding how model size impacts generalization, potentially guiding future model development.

RANK_REASON Academic paper detailing a new theoretical framework for understanding a machine learning phenomenon. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• grokking
• Hugging Face
• Morris et al.
• Yiding Song

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →