Learning Rate

Home Glossary Item Learning Rate
« Back to Glossary Index

The learning rate is a crucial hyperparameter that determines the step size at which an algorithm proceeds while optimizing a loss function.The learning rate determines the extent to which the weights of our network are modified in response to the loss gradient.The learning rate is adjusting the magnitude of updates that are applied to these weights during the process of training a neural network.


The learning rate plays a significant role in reaching the optimal solution. If set too small, the algorithm will need many updates before reaching the minimum, causing the learning process to be slow. This could lead to an underfitting situation where the training does not progress adequately. On the other hand, if the learning rate is too large, the updates could be too drastic, causing the algorithm to oscillate around the minimum or even diverge, resulting in an overfitting scenario where the model fails to generalize from the training data.

Tuning the learning rate is often a balancing act and is a critical step in training machine learning models. Several strategies have been developed to address this, including learning rate schedules that adjust the rate over time or adaptive learning rates where the rate differs for each parameter. Modern optimization algorithms like Adam or Adagrad also employ mechanisms to adaptively change the learning rate based on the characteristics of the data and the specific iteration in the training process, leading to models that learn more efficiently and effectively.

« Back to Glossary Index