The difference between positive samples, negative samples, simple samples and difficult samples in deep learning (simple and easy to understand)

Catalog

• ​ ​1. Before the order ​​
• ​ ​2. A term is used to explain ​​
• ​ ​3. Illustrate with examples ​​
• ​ ​4. reference ​​

1. Before the order

When reading a paper or reading some blogs , This noun often appears ： Positive sample 、 Negative sample 、 Simple samples and difficult samples , Recently, in order to deepen my understanding of this aspect , Refer to some information on the Internet , Sort out the differences between these , It's convenient for you to view and for beginners to quickly understand .

2. A term is used to explain

Positive sample ： For the target category corresponding to the truth value, the sample is a positive sample .
Negative sample ： This sample is a negative sample for all other target categories that do not correspond to the true value .
Simple sample ： Samples with small error between prediction and truth label .
Difficult samples ： Samples with large error between prediction and truth label .

Under the supplement ： The truth value is actually the value of the sample we chose , For example, the current sample is cats , Other dogs 、 Horses and the like are not really worth it .

3. Illustrate with examples

We use cat dog classification as an example to illustrate what positive and negative samples are ：

label = dog All the pictures are A positive sample of this class , Other pictures as negative samples
label = cat All the pictures are A positive sample of this class , Other pictures as negative samples

Empathy , In the target detection task

box The target object in is the positive sample , No target object , That is the prospect , As a negative sample

Next, we are using image classification to illustrate what is a simple and difficult sample ：

If we need to identify horses 、 sheep 、 There are three categories of cattle .
Give a picture of a horse . For prediction, this sample is Positive sample , For predicting sheep and cattle, the sample is Negative sample .
Truth value one-hot label ：[1, 0, 0], When the probability distribution is predicted to be [0.3, 0.3, 0.4] when ,
Truth value one-hot Labels differ greatly , Then the sample is Difficult samples .
And predicted [0.98, 0.01, 0.01] when , And truth one-hot The label difference is small , Then the sample is Simple sample .

For difficult samples, because the truth value one-hot label [1,0,0] According to our artificial thinking, the probability of this picture of a horse is 1, But the predicted probability is 0.3, This is far from reality , So we call such samples difficult samples .

4. reference

​ ​1. The literature 1​​​ ​2. The literature 2​​