lazy training regime

**Lazy Training Regime** is a **theoretical configuration where neural network weights barely change from their random initialization during training** — the network acts essentially as a linear model in the feature space defined at initialization, as predicted by NTK theory. **What Is Lazy Training?** - **Condition**: Very wide networks with small learning rate and/or large initialization scale. - **Feature Freeze**: The features (hidden representations) remain approximately fixed. Only the output layer's linear combination changes. - **NTK Regime**: This is the regime described by Neural Tangent Kernel theory. - **Kernel Method**: In lazy training, the network is equivalent to kernel regression with the NTK. **Why It Matters** - **Theoretical Clarity**: Lazy training is mathematically tractable — convergence and generalization can be proven. - **Poor Features**: Lazy training doesn't learn features — it relies on random features from initialization. This limits performance. - **Practical**: Real networks that achieve SOTA performance operate in the *feature learning* regime, not lazy training. **Lazy Training** is **the couch potato of neural networks** — barely moving from initialization and relying on random features rather than learned ones.