Learning Rate

The learning rate is intuitive: it determines the step size with which the model adjusts its parameters at each training update. Imagine a model walking across a landscape looking for the valley at the lowest point (the optimal solution). A small learning rate means small steps (safe but slow), a large learning rate large steps (fast but risky).

The right learning rate depends on model architecture, data volume, and optimization method. Classic strategies are learning rate schedules (the learning rate decreases in stages during training), warm-up (the learning rate starts low and rises initially), or adaptive methods like Adam that adjust the rate per parameter automatically.

In the models behind Pimcore's AI functions, modern adaptive optimizers like Adam and AdamW are used. These reduce manual hyperparameter search and make model training more robust. For users, the learning rate is not a configuration topic but part of the pre-configured AI architecture Pimcore ships.