Giters

joe-siyuan-qiao / FewShot-CVPR

Few-Shot Image Recognition by Predicting Parameters from Activations

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Mean activations and parameters not correlated on MNIST ?

debasmitdas opened this issue · comments

commented

Your paper is based on the hypothesis that mean activations correlate well with the last layer output parameters. To verify that I used the MNIST dataset and trained a linear softmax classifier on the 784 dimensional features . Then, tsne and plot the features and weight parameters of each class as shown below -
mnist
The blue represents the features of the 10 classes and the orange represents the weight parameters of each class. The oranges are supposed to be close to the centers of the clusters but they are not. Can you explain the reason ? Did you face similar issue before ?

commented

I got the answers. It had to do with the scaling of the W's.