Hypothesis: gradient descent prefers general circuits - LessWrong

Summary: I discuss a potential mechanistic explanation for why SGD might prefer general circuits for generating model outputs. I use this preference to explain how models can learn to generalize even…

x Hypothesis: gradient descent prefers general circuits — LessWrong Gradient Descent Optimization AI Frontpage 46 Hypothesis: gradient descent prefers general circuits by Quintin Pope 8th Feb 2022 AI Alignment Forum 14 min read 26 46 Ω 20 Summary: I discuss a potential mechanistic explanation for why SGD might prefer general circuits for generating model outputs. I use this preference to explain how models can learn to generalize even after overfitting to near zero training error (i.e., grokking). I also discuss other perspectives on grokking and deep learning generalization. Additionally, I d

related reading

• [2201.02177] Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetsarxiv.org
• Just Ask for Generalization | Eric Jangevjang.com
• Do Machine Learning Models Memorize or Generalize?pair.withgoogle.com
• Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetsmathai-iclr.github.io
• Circuit Tracing: Revealing Computational Graphs in Language Modelstransformer-circuits.pub
• Understanding deep learning requires rethinking generalizationarxiv.org
• The Scaling Hypothesis · Gwern.netgwern.net
• [2605.01172] A Theory of Generalization in Deep Learningarxiv.org
• Clare Lyle | What's grokking good for?clarelyle.com
• To Understand Language is to Understand Generalization | Eric Jangevjang.com
• Ambiguous out-of-distribution generalization on an algorithmic task — LessWronglesswrong.com
• What if not Circuits? — LessWronglesswrong.com