Inception-v1_Going_Deeper_With_Convolutions.md

Inception-v1 (Going Deeper With Convolutions)

Status: Read

Author: Christian Szegedy, Wei Liu

Topic: CNNs, CV , Image

Category: Architecture

Conference: CVPR

Year: 2015

Link: https://arxiv.org/abs/1409.4842

Summary: Propose the use of 1x1 conv operations to reduce the number of parameters in a deep and wide CNN

Questions

What did authors try to accomplish?

• The authors try to reduce the number of parameters in a deep CNN whilte still increasing the depth and width of the network

What were the key elements of the approach?

• Use of 1x1 convolutions to reduce the number of parameters
• Stacking of inception modules which combines max pooling, 3x3, 5x5 and 1x1 conv operations together
• Adding intermediate softmax layers to propogate error in earlier layers.

What can you use yourself from this paper?

• Use of 1x1 conv operations to reduce the number of filters and dense connections in a model

What other references to follow?

---