Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 25·8 min read
Human-Centered Learning Mechanics: A Dynamical Framework for Entropy-Regulated Representation Learning
TL;DR
Deep learning training is typically viewed as closed optimization, ignoring real-world constraints like uncertainty and human feedback. Human-Centered Learning Mechanics (HCLM) introduces entropy regularization through information-theoretic dynamics, using geometric entropy surrogates like log-determinant covariance to generate stable optimization forces.
✦ Why It Matters
Engineers can diagnose when entropy regularization fails and select geometric surrogates that produce reliable optimization dynamics in constrained, real-world settings.
Key Takeaways
- Test whether your entropy term generates non-degenerate gradients; weak signals indicate collapse to standard loss minimization.
- Replace softmax-normalized entropy with log-determinant covariance proxies for stronger, more stable optimization forces.
- Use effective entropy framework to diagnose when entropy regularization fails in your training dynamics.
- Apply HCLM to model systems under uncertainty, resource constraints, and distribution shift, not just closed optimization.
Full Summary
Current deep learning theory treats training as isolated optimization, but real systems face uncertainty, resource limits, distribution shifts, and human input. HCLM reframes learning as an open dynamical system using information theory, where entropy regularization only helps when it creates a non-degenerate information force—a directional gradient push—along the optimization path.
Otherwise, entropy terms weaken or collapse into standard loss minimization. The framework introduces effective entropy and studies geometric entropy surrogates: variance-based and log-determinant covariance proxies that approximate information content.
Three main contributions: formalizing entropy regularization via information force and identifying when it fails; deriving convergence and generalization guarantees under explicit assumptions; and interpreting scaling-law behavior as equilibrium between information injection, entropy dissipation, and residual risk. Controlled representation-learning experiments confirm log-determinant covariance entropy induces more stable and powerful information forces than softmax-normalized alternatives.