Optimizer Selection

Optimizer selection

Optimisers differ in how they use gradient history to set the step. The adaptive ones are robust defaults, plain momentum sometimes generalises better, and the choice interacts strongly with learning rate and regularisation.

Method

1. Start with a well-tested adaptive default. It works across a wide range of problems with little tuning, which is why it is the default in nearly every codebase.
2. Understand momentum as smoothing. Accumulating past gradients damps oscillation across narrow valleys and accelerates along consistent directions.
3. Know what adaptivity does. Per-parameter step scaling helps when gradient magnitudes vary widely across parameters, which is typical in deep networks.
4. Decouple weight decay from the gradient. Applying decay as part of the gradient interacts badly with adaptive scaling, and the decoupled variant is the correct default (see regularization-techniques).
5. Tune the learning rate before anything else. It matters more than the optimiser choice by a wide margin (see learning-rate-schedules).
6. Watch the optimiser's memory cost. Adaptive methods store per-parameter state, adding substantially to memory for large models (see gpu-memory-hierarchy).
7. Change one hyperparameter at a time. Optimiser settings interact, and simultaneous changes make results uninterpretable (see experiment-tracking).

Boundaries

Optimiser choice matters less than data quality and architecture for most outcomes. Published settings are tuned to their paper's setup and transfer imperfectly. Second-order methods are rarely practical at scale despite their theoretical appeal.