Different terminologies for ground truth

Question

It's general theory that a supervised learning approach can only get as good an accuracy as the quality of the ground truth. The ground truth is considered the best possible annotation.

However, consider the following scenario: I have roughly annotated ground truth and I'm using a neural network to do some segmentation tasks. Then after training, I use the trained network to do prediction on the entire dataset. Then if there are any false positives that turn out to be actually true positives, then we refine the dataset.

1. Can I call this active learning or would this be called bootstrapping or neither?
2. I was told that what I'm calling "roughly annotated ground truth" isn't ground truth at all and I should be using a different terminology. What exactly should I call it if not ground truth?

Answer 1

1. Can I call this active learning or would this be called bootstrapping or neither?

Neither. Active learning refers to scenarios when the learning algorithm can identify itself samples for which additional annotation is needed. Bootstrapping refers to scenarios when the same model is trained using several different training sets created by sampling with replacement from the original training set.

You use the model predictions for refinement of your "rough ground truth", which is none of the above.

2. I was told that what I'm calling "roughly annotated ground truth" isn't ground truth at all and I should be using a different terminology. What exactly should I call it if not ground truth?

Depends on the context: When talking about the annotations that are used during the supervised learning to train the model, one says "ground truth" regardless of how these annotations were obtained—now they serve as the ground truth that the model is trying to learn.

When talking about the actual dataset, "ground truth" refers to the real, true labels for the samples. In some domains, such as medicine, this is sometimes emphasized as the "golden standard", which means the labels obtained via the best available method (under reasonable conditions). On the contrary, labels obtained via some suboptimal process (e.g. some other automatized algorithm) are called "silver standard". I suppose that would be an appropriate name for your "rough ground truth".

Finally, be aware that if you use your model for refinement of its own ground truth, your accuracy may become biased.