1. Motivation of inception: To make deeper and wider CNN
comes with two drawbacks: limitation of examples and increasing computational resources. ==> need to move from fully connected to sparsely connected architectures.
2. Solutions:
i. use 1 x 1 convolutions before 3 x3 and 5 x 5 ==> reduce redundant infor and reduce dim ( 'keep the network sparse at most places and compress the signals only whenever they have
to be aggregated')
ii. introduce loss functions at hidden layers during training process
a. the model will converge faster (overthrown latter?)
b. loss of final layer: loss of auxiliary classifiers = 1: 0.3
