7
$\begingroup$

What does "ground truth" mean in the context of AI especially in the context of machine learning?

I am a little confused because I have read that the ground truth is the same as a label in supervised learning. And I think that's not quite right. I thought that ground truth refers to a model (or maybe the nature) of a problem. I always considered it as something philosophical (and that's what also the vocabulary 'ground truth' implies), because in ML we often don't build a describing model of the problem (like in classical mechanics) but rather some sort of a simulator that behaves like it is a describing model. That's what we/I call sometimes black box.

What is the correct understanding?

Improve this question
$\endgroup$
1

2 Answers 2

Reset to default
7
$\begingroup$

In the context of ML, ground truth refers to information provided by direct observation (empirical evidence). If you're training an algorithm to classify your data, then the ground truth will be the actual, true labels which could for example be manually annotated by an domain expert. Please note that the models prediction or the inferred labels, are not considered ground truth.

Improve this answer
$\endgroup$
0
5
$\begingroup$

It really depends on what words you put after "ground truth". Sometimes people will talk about "ground truth labels", for example in the context of classification or regression problems. The "ground truth labels" in such a case would refer to the true labels of instances; the labels that we use as target labels for instances from a training set, or the labels that we expect our models to output (and "punish" them for if they fail to do so) when evaluating/testing a trained model. This basically follows razvanc92's answer.

"Ground truth" can also refer to something more abstract though, something that we know exists in some form or another, but we may not even know how to express it. For example, there may be "ground truth laws of physics", the laws of physics that our world "follows". We may build or train a simulator trying to approximate those ground truth functions / laws, but we may not actually know how to explicitly express all of them.

Improve this answer
$\endgroup$
1
  • 2
    $\begingroup$ For the latter consideration, consider also a statistical distribution that generated some training data. You might not be able to observe this underlying distribution directly, but at least you can learn to approximate it. $\endgroup$
    – Daniel B.
    Sep 29, 2020 at 17:14

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .