Small sample learning - getting started

The term

Support Set / Query Set and  N-way k-shot

Traditional image classification

Small sample learning

The effect of small sample learning ：

• Give the model a query sample （query set）, This sample belongs to a new , Classes you haven't seen before ,
• Give it a support set S（support set）, The model must use the information from the support set to learn to query set To classify . The support set consists of n One from k Sample composition of different invisible classes , This is it. N-way k-shot

Be careful N and k By Testing data Decisive , And Training data irrelevant ;

How to start from the original training dataset and testing dataset Get these Support set and Query set Well ？

  In the box on the right testing data It can also mean this ： 

training Will provide a large number of sample , Make model learn to learn.training The class of procedure is called base class.

testing There are only a few samples in the process , test model few shot learning The ability of .testing It contains training A new class that I haven't seen before （novel class）

episode

stay Matching networks In this paper, the definition of training cycle is put forward episodes The concept of , To distinguish big data training epochs, stay episodes In the cycle , All for the service of few-shot Subcategory sample training of tasks , This subcategory is different from epochs Training of sub samples in all categories ,. quite a lot meta-learning I also like to use episodes This word , And the corresponding neural network mini-batch It's more appropriate .

give an example ：20-way 5-shot problem ：

• Every episodes extracted 20 Categories , Not all categories （ Tradition softmax Classification training should load all categories ）;
• Each category has 5 individual examples Available for training , Because there are still points in training support-sets+query-sets,5-shots The scenario needs at least 5+1 A sample , At least one query example Where to support-sets Example of distance （ classification ） Judge .
• The test set can be transformed into 5-way 1-shot And so on to evaluate the generalization ability of the model .
• Last of all epoch Also set how many episodes To form a , It seems that everyone likes to use 1000 As a typical value .

domain gap / catagory gap

Used to describe data set differences

 domain gap： Here's the picture , It may be the same kind , But clarity , The light is different

 category gap： Different kinds

Reference material

machine learning -47-ML-03-Metric-based Approach & Train+Test as RNN( Meta learning -support set and query set Methods for the same network )_ The fog will always solve the blog -CSDN Blog