Microsoft Azure AI Fundamentals (AI-900)Describe fundamental principles of machine learning on AzureHard
A data scientist is training a deep learning model for image recognition. During training, the model's accuracy on the training data consistently improves, reaching nearly 100%, but its accuracy on a separate validation dataset starts to decrease after a certain number of epochs. What is the most likely issue occurring?
- AUnderfitting
- BOverfitting
- CInsufficient data
- DData leakage
Show answer & explanationAnswer & explanation
Correct answer: B. Overfitting
This classic symptom describes overfitting. The model has learned the training data too well, including its noise and specific patterns, and is failing to generalize to new, unseen data (the validation set). The decreasing validation accuracy while training accuracy continues to rise is a strong indicator.
Why the other options are wrong
- A. Underfitting occurs when the model is too simple to capture the underlying patterns, leading to low accuracy on both training and validation data.
- C. Insufficient data might lead to poor performance overall, but the specific pattern of high training accuracy and decreasing validation accuracy points more strongly to overfitting.
- D. Data leakage would typically lead to an *overly optimistic* validation accuracy, as the model would have indirectly seen information from the validation set during training.
Overfitting
A phenomenon where a machine learning model learns the training data too precisely, including noise and specific patterns, leading to excellent performance on training data but poor generalization to new, unseen data.
- High training accuracy, low validation/test accuracy.
- Caused by overly complex models or insufficient training data.
- Mitigation techniques include regularization, early stopping, cross-validation, more data, feature selection.
Memory trick: Overfitting is 'Over-complicated' and 'Over-confident' on training data.