RESNET Architecture:
Prerequisites:
VGGNet OR VGG16 : VGG16 Architecture
DEVELOPED BY MICROSOFT ON 2015
PROBLEM OBSERVED: Regular Neural Networks has a problem that as number of layers increase, even after many epochs the train error and test error of 56 layered Neural Network is worse than the 20 layered Neural Network.
Inorder to solve the above problem, The Microsoft
team has come up with an idea known as residual unit or identity unit which
have skip connection.
Here the key point is during regularization, if we have unimportant layers then we can avoid them with skip connection. so that there will be no impact on our network.
If we observe the above figure, we have two layers, If those both layers are useless then during regularization the weights become close to zero and it will effects the overall architecture. To avoid this kind of situation the skip connection has provided so that eventhough the layers are not useful we can able to skip and transfer the data efficiently to other layers.
Wow, Great Innovation.
Advantages:
1. Adding layers wouldn't hurt performance as regularization will skip over them.
2. We can strongly believe that as layers increases performance will increase.
Disadvantages:
1. Increased complexity in the Architecture.
Comments
Post a Comment