Abstract: Few-shot learning aims to build a classifier that recognizes new unseen classes given only a few samples. Existing traditional metric learning methods map the samples to the shared embedding space, and calculate the feature similarity in this space for classification, but only map the features of samples independently while neglecting to observe the whole task. At the same time, the basic prototypes computed in the low-data regime are biased against the expected prototypes, resulting in low generalization on the query set. In view of the above problems, a feature relation dependent network is proposed (FRDN). The feature relation dependent network consists of two modules: Firstly, the relation mining module can fully mine the intra-class and inter-class relations in the task, use it as the self-attention values to adjust the class clusters to obtain a more discriminative task-adaptive embedding spaceand calculate basic prototypes; Then, the bias diminishing module is used to correct the initial prototype to obtain an optimized prototype with higher generalization on the query set, further improve the classification accuracy. On the MiniImagenet dataset, the 1-shot accuracy of the method is 59.17%, and the 5-shot accuracy is74.11%, which are 6.13% and 2.83% higher than that of the traditional metric learning method; on the CUB dataset, increases of 9.3% and 2.74% are reached respectively.