undertrained

Undertrained is a term used to describe models or systems that have not been trained sufficiently to learn the patterns in their data, resulting in suboptimal performance. In machine learning, undertraining often leads to underfitting: the model cannot capture the underlying structure, producing simple or biased predictions on both training and new data.

Causes of undertraining include insufficient data, too few training iterations or epochs, an inappropriate learning rate, excessive regularization, and limited model capacity. Poor data quality, noisy labels, or unrepresentative training sets can also hinder learning and leave the model undertrained even after long training.

Indicators of undertraining include high training error, low or stagnant validation accuracy, and predictions that fail to differentiate between inputs. The consequence is poor generalization and consistently weak performance across tasks, especially on unseen data.

Mitigation strategies involve increasing training data or data quality, allowing more training time, adjusting model capacity (for example, adding layers or units), and tuning hyperparameters such as the learning rate and batch size. Reducing regularization if it is too strong can help, as can data augmentation, transfer learning, and cross-validation to guide parameter selection. Monitoring learning curves is important to detect undertraining early and distinguish it from other issues such as overfitting.

Outside machine learning, the term can describe people or systems that have not received sufficient training, leading to lower readiness or competence in their roles. See also underfitting, learning curves, data augmentation, and transfer learning.