Outlier detection and open set identification (1)

---

Preface

   This section introduces the definition of open set recognition and its difference from closed set recognition . I hope you can learn something ~

---

One 、 What is open set recognition ？

   Open set recognition , english Open Set Recognition, Shorthand for OSR.

   background ： Machine learning models trained in a closed world usually incorrectly classify test samples belonging to unknown categories into known categories with high confidence . Some literatures call this overconfident behavior of the model “ The arrogance of the model ”. therefore ,OSR On 2013 Was proposed in , To solve this problem .

   Definition ：“ Open set recognition ” Multiple classifiers are required to meet the following two requirements at the same time ：

• Yes, the test set belongs to “ Known categories “ Accurately classify the pictures of ;“ Known categories ” Represents the categories that exist in the training set .
• Detected ” Unknown ” Category , “ Unknown category ” Does not belong to any category in the training set .

In summary, it is , Identify data belonging to known classes as specific categories , Recognize data that does not belong to a known class as an unknown class , That is, exception class .

Two 、 Closed set identification VS Open set recognition

   In the paper 《Open Set Classification for Signal Diagnosis of Machinery Sensor in Industrial Environment》 The difference between open set recognition and closed set recognition is given in ：

On the left is closed set recognition , On the right is open set recognition .

   Under the closed set assumption , The algorithm is based on the existing samples in the training set , Divide the corresponding space for each class . Pictured 3(a) Shown , The dotted line represents the unbounded decision boundary of each class . In the test phase , Samples will be allocated to one of the spaces , But data from unknown categories may be incorrectly classified into a known category （ As shown by the red star symbol in the figure ）.
   In an open set scenario , Based on the assumption that some classes in the training set are unknown , The algorithm determines a finite region associated with each class . As shown in the figure , The decision boundary expressed by solid lines is bounded . If a sample is located in a region of a known class , Then it will be identified as this class . On the other hand , If it is in a space that is not associated with any known class , Then it will be rejected as an unknown class . under these circumstances , The algorithm can be seen as n+1 classifier , among n Indicates the number of known classes in the training stage ,1 Represents an unknown class .

   In short , Closed set recognition is based on known classification , There will be no other results , Open set recognition can produce a new class , And detect the abnormality .

summary

   The above is the definition of open set recognition , I hope you can get something ~
   The reference website is as follows ：
   stay away from crowds ？ abnormal ？ A new class ？ The opening episode ？ Out of distribution detection ？ Understand the similarities and differences in one article ！