研究一门科学,就要先溯本求原,学习它的发展历史,然后再了解其研究热点。

name介绍时间论文
LeNet第一个成功应用于mnist数据集的卷积神经网络1988Gradient-Based Learning Applied to Document Recognition
AlexNet提出ReLU、Dropout、最大池化、LRN、GPU加速、数据增强2012ImageNet Classification with Deep Convolutional Neural Networks
NIN提出mlpconv层(1×1卷积+mlp),提出全局平均池化替代全连接层2014Network In Network
ZF使用反卷积对CNN的中间特征图进行可视化分析2013Visualizing and Understanding Convolutional Networks
Overfeat采用offset池化2013OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks
VGG3×3卷积核,2×2池化核,训练阶段最后的三个FC,测试时换成conv2014Very Deep Convolutional Networks for Large-Scale Image Recognition
GoogleLeNet核心组件为inception,全局平均池化替代全连接层2014Going Deeper with Convolutions(Inception-V1)
Inception-V3并联不同尺寸卷积核的处理过程2015Rethinking the Inception Architecture for Computer Vision
FCN逐像素预测,解决语义分割问题2015Fully Convolutional Networks for Semantic Segmentation
ResNet允许原始输入信息直接传到后面的层2015Deep Residual Learning for Image Recognition
DenseNet加强特征复用,缓解梯度消失2017Densely Connected Convolutional Networks
DPN融合ResNeXt和DenseNet的核心思想2017Dual Path Networks
RAN残差注意力网络是一种应用了注意力机制的卷积神经网络,可将最先进的前馈网络架构融合到端到端的训练过程中。2017Residual Attention Network for Image Classification
SE-Net包含特征压缩、激发(特征通道权重计算)和重配权重等过程2018Squeeze-and-Excitation Networks
Logo

北京人形旗下天工造物具身智能开源社区,聚焦具身天工与慧思开物两大平台

更多推荐