NorbertZheng / read-papers

My paper reading notes.
MIT License
8 stars 0 forks source link

Sik-Ho Tang | Review: RiR -- ResNet in ResNet (Image Classification). #103

Closed NorbertZheng closed 1 year ago

NorbertZheng commented 1 year ago

Sik-Ho Tang. Review: RiR — ResNet in ResNet (Image Classification).

NorbertZheng commented 1 year ago

Overview

In this story, RiR (ResNet in ResNet) is briefly reviewed. RiR attempts to generalize the ResNet block for Residual Network by

so as to improve the accuracy. And it is published in 2016 arXiv with about 80 citations. I treat RiR as a kind of sidetrack papers for ResNet.

RiR is designed in the way that the input signal can go through the network in between ResNet and standard ConvNet.

NorbertZheng commented 1 year ago

ResNet in ResNet

A generalized residual architecture is introduced to combines residual networks and standard convolutional networks in parallel residual and non-residual streams. image (a) 2-layer ResNet block. (b) 2 generalized residual blocks (ResNet Init). (c) 2-layer ResNet block from 2 generalized residual blocks (grayed-out connections are 0). (d) 2-layer RiR block.

Below table summarized the conditions when residual and transient connections are on/off: image The 18-layer + wide RiR used for experiments: image 18-layer + wide RiR.

NorbertZheng commented 1 year ago

Results

CIFAR-10

image CIFAR-10.

CIFAR-100

image CIFAR-100.

NorbertZheng commented 1 year ago

Further Analyses

Effect of Zeroing Learned Connections

image Accuracy Change When Zeroing Learned Connections of One Layer in Each Stream.

Adding More Layers Within a Residual Block

image Accuracy When Adding More Layers Within a Residual Block.

image Accuracy of RiR and ResNet with different numbers of layers per block on CIFAR-10.

Study on Number of Blocks and Layers Per Block

image Number of Blocks and Layers Per Block.

NorbertZheng commented 1 year ago

Besides RiR, there is also another sidetrack paper, ResNet of ResNet (RoR), firstly appeared in 2016 arXiv, and recently published in 2018 TCSVT. Hope I can cover it in the future.

NorbertZheng commented 1 year ago

References