3
$\begingroup$

My background knowledge: Basically, supervised learning is based on labeled data. Using the labeled data, the machine can study and determine results for unlabeled data. To do that, for example, if we handle picture issue, manpower is essentially needed to cut raw photo, label on the photos, and scan on the server for fundamental labeled data.

I know it sounds weird, but i'm just curious if there are any algorithms/system to make a label automatically for supervised learning.

$\endgroup$
1
  • $\begingroup$ If we could automatically make the labels then we would already have our algorithm. $\endgroup$
    – kbrose
    Commented Apr 27, 2018 at 14:04

2 Answers 2

1
$\begingroup$

The problem you ask is probably Semi-supervised learning.

In a Semi-supervised learning problem, you have a small amount of labeled data with a large amount of unlabeled data. The goal of this problem is to build a model that has higher accuracy than using labeled data alone.

On the other hand, if you have no labeled data at the beginning, then you might have to use some automatic system (e.x: sensors) to determine the label of your data. In this paper about robotic grasping, Sergey Levine automatically detects an attempt of grasp as success or failure based on the openness of the grippers (aka fingers) of the robotic arm and the difference of images before and after the attempt. He called this type of problem Self-supervised learning.

Hope this help.

$\endgroup$
1
$\begingroup$

There are a number of algorithms that can assist in the creation of labeled datasets. But if you're talking about supervised learning you must aquire labels one way or another.

Active Learning is a strategy where you start with a blank model and unlabelled data. The model then selects some examples and asks for annotations, over time the model starts to get more confident and only asks for labels for the examples it has low confidence in.

Weak Supervision is another method where instead of manually labeling the data, you write label functions that suggest labels, one example is simple dictionary matching (distance learning) but label functions can be any code. A label model is used to combine the label suggestions.

More recently, One Shot or Few Shot learning looks to learn from only 1 or a few examples per class.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.