Graph Structure of Neural Networks何凯明团队

explore the implications of graphs and how graph theory tools can helps us to learn better presentation.
before jumping to deep learning ,i want to say that graphs are everywhere in real world .
首先，从图结构角度来回顾一些神经网络。

the first one is original LeNet， the connections between the s2 and c3 layers are actually not fully connected . In the LeNet , the handcrafted connectivity pattern(模式) was used. we can see the adjacency（邻接）matrix on the right ,so according from the paper , the motivation of doing this is for different feature maps are forced to extract different hopefully complementary features because they get different set of inputs, however there’s no experimental justifications whether this pattern