The history of CNN structure evolution shown in the following figure begins with a neurocognitive model and a convolutional structure has emerged, but the first CNN model was born in 1989 and LeNet was born in 1998. With the introduction of ReLU and dropout, and the historical opportunities brought by GPU and big data, CNN has ushered in a historical breakthrough in 12 years. After 12 years, CNN's evolutionary path can be summarized into four: 1) deeper network, 2) enhanced convolution mode function and convergence of appeals, 3) from classification to detection, and 4) addition of new functional modules.
Start - LeNet
In 1998, LeCun proposed LeNet and successfully applied it to American handwritten digit recognition. But soon, CNN's edge was overshadowed by SVM and hand-designed local features.
Turning point - AlexNet
The reason why AlexNet is successful, deep learning can return to the historical stage, because:
Nonlinear activation function: ReLU
Ways to prevent overfitting: Dropout, Data augmentaTIon
Big Data Training: Million Level ImageNet Image Data
Other: GPU implementation, use of LRN normalization layer
The first evolution path: the network becomes darker
VGGNet can be seen as a deeper version of AlexNet, see Karen Simonyan and Andrew Zisserman's paper "Very Deep ConvoluTIonal Networks for Large-Scale Visual RecogniTIon".
VGGNet and GoogLeNet mentioned below are the second and first place in the 2014 ImageNet competition, with Top-5 error rates of 7.32% and 6.66% respectively. VGGNet is also a five convolution group, a 2-layer fully-connected image feature, and a 1-layer fully-connected classification feature, which can be regarded as a total of eight parts like AlexNet.
The second evolutionary path: enhanced convolution module
First, let's talk about the idea of ​​NIN (Network in Network) (see Min Lin and Qiang Chen and Shuicheng Yan's paper "Network In Network"), which makes two improvements to the traditional convolution method: the original linear volume The linear convoluTIon layer becomes a multilayer perceptron; the improvement of the fully connected layer is a global average pooling.
MIN caused the evolution of the convolutional neural network to another evolutionary branch, the function of the enhanced convolution module. In 2014, GoogLeNet (Inception V1) was born. Google's GoogLeNet is the winner of the 2014 ILSVRC Challenge, which reduced the Top-5 error rate to 6.67%. For more on GoogLeNet, see the paper "Going Deeper with Convolutions" by Christian Szegedy and Wei Liu.
ResNet is still: not the deepest, only deeper (152 layers). I heard that the current number of layers has exceeded one thousand. ResNet's main innovation is in the residual network, which is one of the hottest AlphaGo Zero technologies. As shown in Figure 11, in fact, the proposal of this network is essentially to solve the problem that it cannot be trained when the level is deep. This kind of network borrowing from the Highway Network idea is equivalent to opening a channel next to the input so that the input can go directly to the output, and the optimized target changes from the original fitted output H(x) to the difference between the output and the input H(x)-x. Where H(X) is the original expected mapping output of a layer and x is the input.
Considering that some telecommunication are built in remote areas, our engineers configure our 48V 200Ah lithium iron phosphate (LiFePO4) battery with GPRS data transmission unit which improve the convenience of battery monitoring. Meanwhile, the GPS function is an anti-theft feature offering battery tracking, which is other guarantee for the battery safety. UFO 48V 200Ah Telecom Battery can give you confidence in good and safe performance of Lithium Battery for telecom tower.
Feature of Rack Mount LiFePO4 Battery for Telecom Towers
â— Safe LiFePO4 Battery Cells
â— Modular Design for Limited Space
â— GPRS for Battery Remote Monitoring
â— Communication Function
â— Flexible capacity scalability
Protection of Rack Mount LiFePO4 Battery for Telecom Towers
â— Over charge/ discharge Protection
â— Over current Protection
â— Over voltage Protection
â— Over temperature Protection
â— Short Circuit Protection
High Efficiency Battery,Robot Battery Specification,Robot Lithium Battery Packs,Lithium Iron Phosphate Batteries
ShenZhen UFO Power Technology Co., Ltd. , https://www.ufobattery.com