Not every machine learning problem comes from a model being too complicated. Sometimes the opposite happens: a model is too simple, too undertrained, or too constrained to even capture the basic patterns sitting right in front of it. This problem is called underfitting, and while it gets less attention than its more famous counterpart, overfitting, it is just as capable of ruining a model’s usefulness.
What Is Underfitting?
Underfitting occurs when a machine learning model fails to learn the underlying patterns in the training data well enough to make accurate predictions, even on the data it was trained on. An underfit model is too simplistic for the complexity of the problem it is trying to solve, resulting in poor performance across the board, on training data, validation data, and real-world data alike.
Unlike overfitting, where a model performs great on training data but poorly elsewhere, underfitting is consistent in its mediocrity. An underfit model simply has not captured enough of the real signal in the data to perform well anywhere.
How to Recognize Underfitting
The clearest sign of underfitting is high error, or low accuracy, on both the training set and the validation set. If a model cannot even perform well on the data it directly trained on, it is a strong indication that the model lacks the capacity or the training to capture the patterns present in that data.
On a training curve, underfitting typically shows up as a loss value that plateaus early and stays high, for both training and validation loss together, rather than the widening gap between the two that characterizes overfitting. The model essentially stops improving well before it has learned anything particularly useful.
Why Underfitting Happens
A number of common causes lead to underfitting:
- Model too simple: A linear model applied to a problem with genuinely complex, non-linear relationships will structurally be unable to capture those patterns, no matter how much data it sees.
- Insufficient training: Stopping backpropagation too early, before the weights have had enough iterations to converge on useful values, leaves the model undertrained.
- Too few relevant features: If the input data is missing the information actually needed to make accurate predictions, no amount of model complexity can compensate for that missing signal.
- Excessive regularization: Regularization techniques are meant to prevent overfitting, but applying them too aggressively can overcorrect, restricting the model so heavily that it cannot capture real patterns either.
- Poor feature engineering: Raw data that has not been properly cleaned, transformed, or structured can obscure patterns that a model would otherwise be able to learn.
A Simple Analogy
Imagine trying to sketch a detailed, realistic portrait using only three thick, blunt crayons and a few minutes of effort. No matter how hard the artist tries, the tools and the time available are simply insufficient to capture the subtlety and detail of a real face. The result will be a crude, oversimplified sketch that misses most of the defining features entirely. That is underfitting: the model, like the crayon sketch, lacks either the capacity or the effort needed to represent the real complexity of what it is trying to depict.
How to Fix Underfitting
Because underfitting generally stems from a model being too constrained, the fixes typically involve giving it more capacity or more opportunity to learn:
- Increase model complexity: Adding more layers, more parameters, or switching to a more sophisticated architecture gives the model greater capacity to capture complex patterns.
- Train for longer: Allowing more training epochs gives the weights more opportunity to converge toward values that genuinely minimize the loss function.
- Add more relevant features: Enriching the input data with additional, meaningful variables can give the model the signal it needs to make better predictions.
- Reduce regularization: Dialing back overly aggressive regularization penalties can free the model to fit the real patterns in the data more closely.
- Improve feature engineering: Better data preprocessing, transformation, and selection can expose patterns that were previously hidden from the model.
Underfitting and Overfitting: Two Sides of the Same Problem
Underfitting and overfitting represent opposite failure modes along the same spectrum. An underfit model is too rigid and misses real patterns; an overfit model is too flexible and chases noise as if it were signal. The objective of training a good machine learning model is to land somewhere between these two extremes, often described as the bias-variance tradeoff. Too much bias, and the model is underfit; too much variance, and it is overfit. Finding that balance is less about following a fixed formula and more about iterating: adjusting model complexity, training duration, and data quality while watching how performance shifts on a held-out validation set.
Why It Matters
Underfitting might sound like a less dramatic problem than overfitting, since it does not create the illusion of a great model that quietly fails later. But that straightforwardness has a cost of its own: an underfit model can be dismissed as simply “not working,” leading teams to abandon a promising approach or dataset when, in reality, a more capable model or more complete training would have revealed the patterns all along.
Recognizing underfitting for what it is, rather than mistaking it for a dead end, is often the difference between giving up on a machine learning approach prematurely and finding the right adjustments that unlock genuinely useful performance.