Hyperparameter tuning is the process of finding the best configuration settings for a model's training process. Unlike model weights (which are learned from data), hyperparameters are set before training begins and control how learning happens: learning rate, batch size, number of epochs, optimizer choice, weight decay, augmentation strength, architecture-specific settings (anchor sizes, confidence thresholds), and many more.

The simplest approach is manual tuning based on experience and published baselines. Grid search tests every combination of a predefined set of values but scales poorly (10 hyperparameters with 3 values each means 59,049 combinations). Random search samples configurations randomly and consistently outperforms grid search because it explores more of the space per trial. Bayesian optimization (Optuna, Hyperopt, Weights & Biases Sweeps) is smarter: it builds a probabilistic model of which configurations work well and focuses subsequent trials on promising regions.

For computer vision, some hyperparameters matter more than others. Learning rate is almost always the most impactful. Batch size affects both training dynamics and GPU memory usage. Augmentation policy can make or break performance on small datasets. Most practitioners start with the defaults from published model configs (YOLO, EfficientDet, SAM all ship with recommended hyperparameters) and make targeted adjustments based on validation metrics.