Key points
- Underfitting is a model that can't get even its own practice data right. It most often happens when the model is too simple.
- It's easy to spot. When the training score and the test score are low together, that's underfitting. It splits off from overfitting, where only one of the two is low.
- There are four main causes: the model can't represent enough, it wasn't trained long enough, the input it was given is too thin, or the safeguard against drifting was clamped on too hard.
- The fixes run the other way. Make the model bigger, train it longer, add features worth having, and loosen the restraint.
- The right spot sits somewhere between overfitting and underfitting. Training is the work of finding it.
The analogy
Line up several sample bowls in front of a potter's wheel and ask for exact copies: a rice bowl, a plate, a tall-necked bottle, each shaped completely differently. But rest just one flat palm on the clay and spin the wheel, and no matter which sample you point to, out comes the same shape — a plain cylinder. Point at the plate, get a cylinder. Point at the bottle, still a cylinder. One palm pressing down can't shape a foot, a neck, and a rim as separate things. Staring harder at the sample doesn't help either — the range of shapes a single palm can produce is simply too narrow. Underfitting is this state: too narrow a range of shapes to follow even the samples sitting right there in practice.
In detail
It misses what it practiced on first
The hallmark of underfitting is a poor score even on the practice data itself. The samples are sitting right there while it shapes the clay, and the copy still doesn't look like them.
That matters because it narrows down the cause. If it's already missing on the practice data, gathering more data of the same kind won't help — look first at the model's breadth, the training time, and the features it was given.
Overfitting, sitting on the other side, gets almost everything right on the practice data and only falls apart on data it has never seen. The two states look completely different on the surface, so once you know which one you're looking at, you know where to work.
Why it happens
First, the range of shapes it can make is too narrow. That's the difference between one flat palm and fingers working independently. When the layers are too shallow or there aren't enough units inside, there's simply no way to capture the curves hiding in the data.
Second, it wasn't spun for long enough. That's like turning the wheel just a few times and stopping. This one is the easiest cause to fix, since time takes care of it.
Third, the material handed over is too thin. Give it the bowl's diameter and height but never mention whether it has a foot, and no matter how skilled the hand, it can't shape a footed bowl differently from one without. When something the decision actually needs is missing entirely, making the model bigger doesn't help.
One more cause: clamping down too hard on a device meant to stop things from drifting can bring on underfitting too. It's like tying the arms down to stop a shaky hand — the motion that was actually needed gets blocked along with the shake.
More hands change everything
The moment fingers start working independently, the result changes at once. One finger pushes the inner wall while another braces the outside, and a foot appears; narrow the top and a neck appears. Different samples start producing different bowls.
Models work the same way. Add layers, add units inside, and the curves it can capture grow with them. A falling training loss is the signal that it's working.
Push past that point without stopping, though, and it can easily swing to the other side. Copying every fingerprint and speck of clay dust off the sample's surface — that's overfitting. The right spot sits somewhere between the two extremes.
How to spot it
Plot the training loss and the validation loss together as training runs, and it shows up right away. If both lines stall out high and refuse to fall further, that's underfitting. If only the training line keeps falling while the validation line turns and climbs back up, that's overfitting.
Starting out underfit on purpose is actually a useful trick. Run a small model first and see how far the training loss falls, then grow it step by step and watch for the point where the validation loss starts to pull away. Start with a big model right out of the gate, and it's much harder to pin down what's actually holding it back.
More precisely
Underfitting describes a state where a model's capacity falls short of the structure in the data, so both training loss and validation loss stay high. Split along bias and variance, this is the high-bias side; overfitting is the high-variance side.
The analogy breaks down in one place. On the wheel, adding hands changes the shape right away, but a model doesn't necessarily improve just because it's made bigger. The extra size needs matching data and training time to back it up, and without that, the training loss can fall while real performance stays exactly where it was. Some problems also have no answer the data itself can settle, no matter how many hands you add — the loss on data it has never seen won't drop below a certain floor. Pushing the training loss below that floor means the noise has been memorized, which is overfitting. Failing to clear that floor on unseen data isn't called underfitting; what's left over there simply belongs to the data itself. Since there's no way to know that floor in advance, you need a procedure of starting small and growing the model step by step, checking whether the loss keeps falling.
Try it yourself
Common misconceptions
It's easy to think underfitting comes from not having enough data, but actually piling on more examples rarely fixes it. The cause is a model that's too narrow, under-trained, missing needed features, or held back by too tight a restraint.
It's easy to think making a model bigger always solves it, but actually if something the decision needs is missing, or a restraint is clamped too tight, size makes no difference either way.
It's easy to think any low score means underfitting, but actually a high training score paired with only a low test score is the opposite — overfitting — and calls for the opposite fix.
One-line summary
Underfitting is failing to follow even the practice data, and it's fixed by growing the model, training it longer, handing over better material, or loosening the restraint.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-17