generalization-theory

What problem does it solve?

Generalization-theory helps you determine why a model’s training loss is improving while test performance stalls or worsens, distinguishing memorization-driven behavior from genuine signal learning.

Core Features & Use Cases

• Signal/noise diagnosis (eNTK partitioning): classifies training dynamics into coherent “signal” directions versus trapped “noise” residuals to explain overfitting, grokking, and double descent.
• Unified interpretation of training phenomena: maps benign overfitting, implicit bias, and grokking to the same empirical partition geometry instead of treating them as unrelated quirks.
• Intervention selection ladder: recommends the lightest effective change among data, architecture, and optimizer surfaces, including an SNR-style preconditioning strategy.

Quick Start

Use generalization-theory to analyze your run checkpoints and telemetry to decide whether to intervene through data cleanup, architectural bias, or an SNR-aware optimizer adjustment.