Learning Curve
A line showing how the score changes over training
- A learning curve is a line that connects the score at every round of training as it changes.
- The point of watching it isn't to check the current score — it's to decide whether to keep going or stop.
- Still steep means there's still room to improve; flattened out means further rounds won't gain much.
- If the line bends and starts getting worse, training has gone too far. Roll back to whichever round scored best.
- A line that's flat from the start, or jagged throughout, is a sign the setup is wrong, not a sign of the actual score.
Contents
1The analogy
Sit down in a stylist's chair and the stylist never cuts it all off in one pass. A little gets trimmed, a look in the mirror, a little more gets trimmed, another look. Plot a point every time for how close the shape has gotten to what was wanted, and connect the dots — that gives you a line.
The first few snips shape things up fast. After that, each snip barely shows, and past a certain point, it goes too short and actually moves away from what was wanted. What the stylist watches isn't the number of snips taken — it's the shape of this line. Still steep, keep cutting; gone flat, put the scissors down. A learning curve is exactly this line.
2In detail
What goes on the horizontal axis
The most common learning curve puts the training round on the horizontal axis. Score gets measured and plotted after every full pass through the data, and the dots get connected. This line answers the question: "is it still worth continuing right now?"
There's another version that puts the amount of data used on the horizontal axis instead — plotting how the score changes at 100, 500, and 2,000 examples. This line answers a different question: "is it worth collecting more data?" If the line is still climbing, collecting more pays off; if it's already gone flat, ten times the data won't move it much.
How to read the shape
Read the shape by watching the slope, not the height. A steep stretch means there's still more to learn. Stop there and you've walked away without pulling out the full potential.
Once the slope goes gentle, the effort and the payoff have leveled out — doubling the number of rounds from here barely nudges the score. This is where the question of whether to stop starts to matter.
If the line bends and starts dropping, it's already gone too far — like cutting more hair than was needed. The move at that point isn't to keep going, it's to roll back to whichever round scored best.
Turning "when to stop" into a rule
Deciding "that's about enough" by eye means everyone reads it differently. So a rule gets set in advance. The most common one is simply waiting a set number of rounds. Whichever round posted the best score gets remembered, and if that record hasn't been beaten after a set number of further rounds, training stops.
There's a reason for waiting a few rounds instead of stopping the moment it dips. The line isn't perfectly smooth — one or two rounds can dip by chance. Give up the moment it dips once, and the improvement that would have come right after gets missed. Set the wait too long and time gets wasted; too short and training ends early — so this wait count itself gets treated as a value chosen ahead of time.
What a strange shape is telling you
If the score barely moves at all from the very start, training hasn't really gotten going. A step size that's too small, data that isn't feeding in correctly, or a model too simple for the problem — any of these produce a line like this.
A line that swings sharply up and down means the step size is too large, or the chunk of data seen per step is too small. Sometimes the line stays flat for a long stretch and then drops in a sudden step instead — that shape is a sign it struggled early on before finding its footing, and it's worth waiting a bit longer. The height of the line tells you the score; the shape of the line tells you what needs fixing.
3More precisely
A learning curve is a chart with training round or amount of data used on the horizontal axis, and an evaluation metric like loss or accuracy on the vertical axis. Lower is better when the vertical axis is loss, and higher is better when it's accuracy, so check what the vertical axis actually is before reading the shape. It's standard to plot the value measured on training data and the value measured on held-back data together, and it's the held-back one that gets used to decide when to stop.
The comparison breaks down in places too. Cut hair can't be undone, but training can save its state at every round. So even after going too far, the snapshot from whichever round scored best can be pulled back out and used. And a real curve isn't as smooth as a reflection in a mirror — the order the data got shuffled in, plus randomness, always mixes ripples into it, so it needs to be read as a trend across several points, not the up-or-down of any single one. A haircut also has one obvious finish line, a shape everyone agrees looks done, while a learning curve rarely offers a single clean stopping point — the decision is almost always a judgment call made against a rule set in advance.
4Try it yourself
5Common misconceptions
It's easy to think the line dropping steadily means training should keep going, but actually you need to check whether the score on held-back data is improving right alongside it.
It's easy to think one bad point means training has failed, but actually there's always some ripple in the data, so it's worth watching a few more rounds before deciding.
It's easy to think a flat line means the model has hit its ceiling, but actually the step size or the data can be the real problem, and either one produces the same flat line.
7One-line summary
In shortA learning curve is a line connecting the score at every round, and it's read by its slope and shape — not its height — to decide when to stop and what to fix.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02