NeurIPS 2025: Training Untrainable Networks via Alignment

phillip_isola · x · 2026-08-26

MIT researchers propose 'Guidance' to train traditionally untrainable networks by using inductive biases from a guide network via representational alignment. It prevents FCN overfitting, narrows the RNN-Transformer gap, and boosts plain CNN accuracy.

Original post →

More from Research

Research channel →