linear scaling rule

**Linear scaling rule** is the **heuristic that increases learning rate proportionally with batch-size growth** - it is a common starting point for large-batch training but must be validated with stability controls. **What Is Linear scaling rule?** - **Definition**: If global batch is multiplied by k, initial learning rate is multiplied by k as first-order adjustment. - **Intuition**: Larger batches reduce gradient noise, allowing larger optimization step sizes in many regimes. - **Applicability**: Works best within bounded scaling ranges and with suitable optimizer settings. - **Failure Cases**: At extreme batch sizes, linear scaling can destabilize training or hurt final quality. **Why Linear scaling rule Matters** - **Practical Baseline**: Provides simple, widely used initialization rule for distributed scaling experiments. - **Tuning Efficiency**: Reduces search space when moving from single-node to multi-node batch sizes. - **Speed Potential**: Correctly scaled LR can preserve convergence speed at higher throughput. - **Knowledge Transfer**: Rule offers shared language across teams for scaling discussions. - **Optimization Discipline**: Encourages structured rather than arbitrary hyperparameter adjustment. **How It Is Used in Practice** - **Baseline Establishment**: Start from well-performing small-batch configuration and apply proportional LR scaling. - **Warmup Integration**: Use gradual LR ramp-up to avoid early divergence at high effective step sizes. - **Validation Sweep**: Run narrow LR sweeps around linear target and choose by time-to-quality outcome. Linear scaling rule is **a useful first-order guide for large-batch optimization** - it accelerates tuning, but robust results still require empirical validation and stability safeguards.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account