After recently read a few one-shot series of papers, for me, is really the first big ah, have patience with the translation software to read each paper five or six times, it can be considered to understand it
This is a first to write about: Matching Networks for One Shot Learning
This article is mainly about matching, for surport set (to be matched collection) category of the training set may never appear
First, network structure
FIG left just above four surport set is a training set, bottom right bird test sample, together form a five task
The main task is to match the particular test sample belongs to which type surport set of S → cS (x) can be represented by the formula for the probability P (y | x, S) wherein S is surport set
After gf used in the text of the first four convolutional network images obtained by LSTM features and enhancing a global link
Then do softmax on cos category from the last match of the test sample
Model structure:
among them
c is calculated from Cosine
For LSTM is actually a special RNNs network, RNNs stop by the information loop operation, to ensure the continued existence of the information, but at the same time brings long dependency problem
LSTM through the gate structure (gate forgotten, input gate, the output of gate) to add or delete the information, thereby solving the problem of long-dependent.
rnns and lstm structure is as follows:
rnns:
lstm：
Overall network structure of the code as follows:
Last experiment: 1, in the first data set omniglot, this data set data set similar to the handwritten mnist
2. OF imgnet dataset from 100 randomly selected data class data set made of a minimgenet
3. Also on the adjusted data set imagenet:
