AiVibeOverfitting and Underfitting
Overfitting occurs when a model learns the noise and peculiarities of its training data, performing well in training but poorly on new data; underfitting occurs when a model is too simple to capture the real pattern.
Overfitting is detected by comparing performance on training data with performance on separate validation data: a widening gap shows the model is memorising rather than generalising. It is more likely with complex models, small or unrepresentative datasets and long training. Underfitting shows as poor performance on both sets. The balance between the two is often described as the bias-variance trade-off.
Overfitting is common in industrial projects because defect and failure examples are rare and datasets often come from a single machine, shift or product variant. A model may then fail when lighting, materials or operating conditions change, so validation data should represent the conditions found in production.
Remedies include more and more varied data, data augmentation, regularisation such as weight decay, dropout, early stopping, simpler models and cross-validation. Data leakage, where information from the test set influences training, can hide overfitting and produce misleading results, so test data must be kept separate until final evaluation.
Key points
- Overfit models perform well in training but poorly on new data
- Detected by a gap between training and validation performance
- Common when industrial datasets are small or unrepresentative
- Countered with more data, augmentation, regularisation and early stopping
Where AiVibe comes in
AiVibe delivers AI and machine learning services, chatbots and virtual assistants with RAG, MCP tools and voice, AI quality management including bias detection and model validation, and the AIMURUGA AI agent, and builds Intel-based edge AI devices using the Intel Distribution of OpenVINO toolkit.