Liang N M , Hu N X . Recurrent convolutional neural network for object recognition[C]// 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, 2015.
1. Overview
1.1. Motivation
- the visual system recurrent connections are abundant in CNNs
- context can only be captured in higher layers, but cannot modulate the activities of units in lower layers responsible for recognizing smaller objects
In this paper, it proposes Recurrent CNN (RCNN)
- can integrate the context information
- unfolded network has multiple path.
- longer path to learn highly complex features
- shorter path to help gradient BP
- experiments on CIFAR-10, CIFAR-100, MNIST and SVHN
1.2. Recurrent Convolutional Layer (RCL)
- u. forward feature
x. recurrent feature
f. ReLU
- g. Local Response Normaliztion (LRN)
2. Experiments
2.1. Baseline
2.2. CIFAR-10
- more iteration in RCNN led to both lower training and testing error
- more iteration in rCNN led to lower training error but higher testing error
- dropout palyed an important role for RCNN
- simply increasing K led to better results