AI Glossary / AI and machine learning / Beginner
Underfitting
Definition
Underfitting occurs when a model fails to capture useful patterns in its training data and consequently performs poorly. It can result from an overly restrictive model, unsuitable inputs, excessive regularization or an inadequate learning process.
What it means
A model should simplify the world, but it can simplify too much. A rule that predicts the same demand every day may ignore an obvious weekly pattern. Even the historical examples it was fitted on remain poorly explained.
Underfitting differs from overfitting, where a model follows the training examples too closely and fails to generalize. Both can produce poor real-world results, so diagnosing only the final score is insufficient.
The term does not mean that every error can be eliminated by making the model larger. Noisy labels, missing information and genuinely unpredictable variation can also limit performance.
For business readers, the practical question is whether the model has learned the available signal. Compare training and development results, inspect errors and check simple baselines before deciding whether the remedy is more complexity, better features or better data.
How it works
Diagnosis begins with a clear metric and baseline. The team checks whether both training and development performance are unsatisfactory and whether the model misses patterns that should be learnable from its inputs.
It then investigates possible causes. A linear relationship may be too restrictive for a curved pattern. Strong regularization may suppress useful variation. A training procedure may stop too early or fail to optimize effectively.
Controlled experiments change one plausible cause while holding the evaluation procedure stable. Adding a relevant feature, adjusting capacity or improving optimization can reveal whether the limitation is recoverable.
The team continues monitoring development performance as fit improves. Removing underfitting can eventually create overfitting, so success is better generalization rather than the lowest possible training error.
Three practical examples
- Demand planning: a constant prediction ignores a repeated difference between weekdays and weekends.
- Product classification: a model using only image brightness cannot distinguish categories whose differences depend on shape.
- Response-time estimation: a model ignores workload level even though waiting time changes substantially with queue size.
Think of it this way
Think of trying to describe a winding road with one straight line. The description is simple, but it misses bends that matter for navigation.
The analogy has a limit. Some observed bends may be measurement noise rather than a stable pattern. A more flexible description should be checked on independent observations so that it captures useful structure rather than every accidental detail.
Worked example
Seeing a missing weekly pattern
Suppose an invented shop sells ten units on each weekday and 30 on each weekend day. Weekly sales total 5 times 10 plus 2 times 30, or 110 units.
A model that always predicts the weekly daily average returns about 15.7 units every day. It overpredicts each weekday by about 5.7 and underpredicts each weekend day by about 14.3.
A model with a weekend indicator can represent the two different levels in this simplified dataset. That demonstrates a relevant missing distinction, not proof that the same rule will forecast every future week perfectly.
Real demand may vary with promotions, closures and noise. The team would test later weeks and compare errors before deciding whether the additional feature provides durable value.
Illustrative case study
A fictional repair service uses a simple duration model based only on equipment age. Errors remain large even on training records. Technicians point out that fault type is strongly related to repair time.
The team adds a reliably available fault-category input and reviews label quality. It compares results on later jobs, rather than immediately replacing the model with a much larger system.
The lesson is that missing information can resemble insufficient model power. Increasing complexity cannot compensate reliably for an essential input that the model never receives.
Common misconceptions and mistakes
- Poor test performance always means overfitting. Underfitting, population change and data problems can also cause it.
- More training always fixes underfitting. An unsuitable representation or restrictive model may remain limited regardless of training duration.
- Zero training error is the goal. A model can fit noise or memorize examples while becoming less useful on new cases.
Business use
Underfitting is useful language when a model is too crude for the business pattern or has not learned effectively. It encourages diagnosis rather than reflexively collecting more of the same data.
Ask for training and development curves, representative error examples and a comparison with simple domain-informed baselines. Evaluate the cost of new features or complexity against the improvement they deliver. Some residual uncertainty may be unavoidable and should be reflected in the workflow.
People also ask
Can a large model underfit?
Yes. Poor optimization, unsuitable inputs or restrictive settings can prevent a large model from learning effectively.
Does more data always help?
More representative data may help, but additional copies of the same limited pattern will not fix a missing feature or inappropriate model form.
Can underfitting and overfitting occur in different parts of a task?
Yes. A model may miss broad structure while fitting noise in some regions, so detailed error analysis matters.
Check your understanding
A model ignores weekend status and misses a stable weekend sales increase even in training data. What is a sensible next investigation?
Answer: Check whether a reliable weekend feature or a suitable model structure captures the pattern, then evaluate on separate periods. More parameters alone are not the only option.
Related terms
Learn next
Read overfitting and learning curve. Compare one model's training and development errors, then identify a specific missed pattern and a controlled experiment that could explain it.
More in AI and machine learning
- Tucker decomposition - Tucker decomposition represents a multiway array, or tensor, using a smaller core tensor and one factor matrix for each mode.
- Ugly duckling theorem - The ugly duckling theorem is a result associated with Satosi Watanabe showing that, if all possible predicates are treated equally in a particular logical...
- Uncertain data - Uncertain data are data whose values, identities, relationships or existence are not known with complete confidence.
- Uniform convergence in probability - Uniform convergence in probability means that the largest error across an entire indexed collection of random quantities becomes small with high probability...
- Univariate feature selection - Univariate feature selection evaluates input features one at a time and selects those that score well under a chosen statistical criterion.
- Universal portfolio algorithm - The universal portfolio algorithm is an online portfolio-selection method introduced by Thomas Cover.
Sources
Google, Machine Learning Glossary, underfitting: https://developers.google.com/machine-learning/glossary#underfitting. Core concept and possible causes consulted. Weekly-demand arithmetic and repair scenario are original.
Previous: Underdamped Langevin dynamics Next: Undersampling