Underfitting happens when a machine learning model does not learn enough from the training data. The model is too simple, poorly trained, or missing important information.
Model does not learn enough
Training score stays low
Test score also stays low
Model performs poorly everywhere
An underfit model is not useful because it cannot even perform well on the data it studied.
Underfitting detection matters because a model with poor training and test performance is not ready for real use.
Training Accuracy — 58%
Test Accuracy — 56%
The model performs poorly on both familiar data and unseen data.
The most common sign of underfitting is a low training score and a low test score.
Performance on data the model learned from.
Performance on unseen data.
If both scores are low, the model is likely too weak or has not learned enough.
Training performance shows how well the model performs on the training dataset. If training performance is low, the model is struggling even with familiar data.
Training Accuracy — 55%
The model correctly predicts only 55% of the training examples.
The model has not learned the training data well.
Test performance shows how well the model performs on new data it did not train on.
Test Accuracy — 54%
The model correctly predicts only 54% of unseen examples.
The model also fails on new data.
Underfitting usually appears as poor performance on both training and test data.
Training Accuracy — 60%
Test Accuracy — 58%
The scores are close, but both are weak. This suggests underfitting.
Underfitting and overfitting are different problems.
Low training performance and low test performance.
Training Accuracy = 58%
Test Accuracy = 56%
High training performance and low test performance.
Training Accuracy = 99%
Test Accuracy = 65%
A good fit means the model performs well on both training and test data. Underfitting means the model performs poorly on both.
Training Accuracy = 91%
Test Accuracy = 89%
Strong performance on both.
Training Accuracy = 58%
Test Accuracy = 57%
Weak performance on both.
Loss measures prediction error. For underfitting, both training loss and test loss are usually high.
Training Loss — 1.20
Test Loss — 1.25
The model makes large errors on both training data and unseen data.
In classification, underfitting can appear as low accuracy, low precision, low recall, low F1 score, or weak ROC-AUC.
Training Accuracy — 57%
Test Accuracy — 55%
Training F1 Score — 0.42
Test F1 Score — 0.40
The model is weak on both training and test data.
In regression, underfitting can appear as high training error and high test error.
Training RMSE — 65
Test RMSE — 68
The model makes large numerical errors on both familiar and unseen data.
Underfitting can happen for many reasons. Detecting the cause helps decide how to fix the model.
A model may underfit if it is too simple for the problem.
Only Feature — Number of bedrooms
Missing Features
The model does not have enough information.
A model may underfit if it has not trained long enough.
Training Accuracy — 45%
Test Accuracy — 44%
Possible fix: Train for more epochs while monitoring validation performance.
If the features are weak, irrelevant, or incomplete, the model may underfit.
Regularization helps prevent overfitting, but too much regularization can cause underfitting.
Model weights are forced too small.
Result — The model cannot capture important relationships.
Possible Fix — Reduce regularization strength.
Some algorithms are not suitable for certain problems.
A simple linear model is used for a complex, non-linear relationship.
Problem — The model cannot represent the real pattern.
Possible Fix — Choose a model that can capture more complex patterns.
Poor data quality can make it difficult for the model to learn useful patterns.
Preprocessing is useful, but too much preprocessing can remove important meaning.
Original Text — "This product is not good."
Poorly Processed Text — "product good"
The meaning changed. The model may learn weaker patterns.
Training curves can help detect underfitting while the model is training.
Epoch 1 — Training 45%, Validation 44%
Epoch 10 — Training 55%, Validation 54%
Epoch 30 — Training 58%, Validation 56%
Learning curves show how performance changes as more training data is added.
Training Score — Low
Validation Score — Low
Gap — Small
The model may be too simple or missing important features.
Sometimes training and test scores are close together, but both are poor. This is not good generalization.
Training Accuracy — 61%
Test Accuracy — 60%
Gap — Small
The model is consistently weak. A useful model needs strong performance, not just similar scores.
Underfitting is often connected to high bias. High bias means the model makes overly simple assumptions about the data.
Oversimplified Assumption — House price only depends on bedrooms.
Reality — House price also depends on location, size, condition, market trends, and more.
The model is too limited.
Underfitting is usually related to high bias, while overfitting is usually related to high variance.
Model is too simple and misses patterns.
Leads to underfitting.
Model is too complex and memorizes training data.
Leads to overfitting.
Only Feature — Number of bedrooms
Missing — Location, square footage, age, bathrooms, nearby schools, market trends
Training RMSE — $120,000
Test RMSE — $125,000
The model performs poorly everywhere.
Only Feature — Email length
Missing — Suspicious links, sender reputation, keywords, formatting, attachments
Training Accuracy — 60%
Test Accuracy — 59%
The model is too simple to identify spam patterns.
Underfit Model — Too few layers or too little training
Training Accuracy — 52%
Test Accuracy — 51%
The model barely performs better than guessing.
If the model is too simple, increasing complexity can help.
Goal: Use a model that can capture the pattern without overfitting.
Adding useful features gives the model more information to learn from.
Better features can reduce underfitting.
Feature engineering can create stronger inputs from raw data.
Raw Feature — Birth year → Engineered Feature — Age
Raw Feature — Purchase date → Engineered Features — Day of week, month, holiday season
Good engineered features can help the model learn patterns more easily.
If the model has not trained enough, more training may help.
Epoch 1 — Training Accuracy = 45%
Epoch 10 — Training Accuracy = 65%
Epoch 30 — Training Accuracy = 82%
Training longer helped the model learn. Monitor validation performance to avoid overfitting.
If regularization is too strong, reducing it may help the model learn more.
Model weights are forced too small.
Result — The model cannot capture important relationships.
Fix — Reduce regularization strength carefully.
Hyperparameters affect how a model learns. Poor hyperparameter choices can cause underfitting.
Tuning them can improve performance.
A model has 57% training accuracy and 55% test accuracy. Explain why this suggests underfitting. Then list three changes that could help the model learn stronger patterns.
Ask the AI if you need help understanding underfitting, low training scores, low test scores, high bias, weak features, training curves, or how to improve an underfit model.