Undertraining

Undertraining is a term used in machine learning to describe a situation in which a model has not learned sufficiently from the training data, leading to poor performance on both the training set and unseen data. It is closely related to underfitting, which is the broader concept of a model being unable to capture the underlying patterns of the data. Undertraining can arise from limited data, insufficient training time, or model limitations that prevent learning.

Causes and signs include too few training epochs, overly aggressive regularization, insufficient model capacity, poor data quality, or a training process that terminates prematurely. Indicators are high bias, slow or no improvement in training accuracy with more epochs, and similar high error on training and validation sets.

Remedies involve increasing model capacity (more layers or units), reducing regularization, increasing training time, tuning learning rate and optimization schedule, improving feature engineering and data quality, and using data augmentation or synthetic data to expand the training set. Also check for data leakage and ensure the training set is representative. Use learning curves to monitor progress and avoid premature stopping.

Notes on terminology: terminology usage varies; many practitioners prefer underfitting to describe the core problem. The term undertraining emphasizes the training process rather than the model’s bias, but both describe related issues.