generalization-theory

When training loss improves but test performance stalls, it indicates memorization-driven behavior rather than genuine signal learning. An empirical neural tangent kernel partition diagnoses this by classifying training dynamics into coherent signal directions versus trapped noise residuals.

eNTK partitioning explains grokking delays and double descent by mapping these training phenomena to the same empirical partition geometry. It distinguishes coherent signal directions from trapped noise residuals, treating them as unified training dynamics rather than unrelated quirks.

The best way to select an intervention is using a selection ladder that recommends the lightest effective change among data, architecture, and optimizer surfaces. This includes applying an SNR-style preconditioning strategy to increase signal-channel SNR and improve generalization.

Yes, you need training-dynamics telemetry from run checkpoints to diagnose memorization versus signal learning. Interpreting this telemetry through coherence classification allows the empirical neural tangent kernel partition to determine if training loss improvements reflect genuine signal learning.