issues
search
hysic
/
NNDL_ReadingNotes
Reading notes for Neural Networks and Deep Learning
0
stars
0
forks
source link
Chp.6 Deep Learning
#6
Open
hysic
opened
7 years ago
hysic
commented
7 years ago
卷积神经网络(CNN)
现代的CNN最早由由Yann LeCun等人在1998年发表的
论文
提出。
三个基本概念
局部感受野(Local receptive fields):感受野(feature map, e.g.5 * 5)步长(stride length),一个卷积层一般有多个feature maps
共享权重和偏差(shared weights and bias):比如5*5的感受野,共享25个权重和1个偏差,构成了一个filter或kernel ,探测某一种特定的feature;共享权重和偏差可以大大减少参数个数,还可以满足平移不变性(translation invariance)
池化(pooling):降维(e.g. 2*2个神经元 -> 1个神经元);max-pooling/L2-pooling;
hysic
commented
7 years ago
CNN 实践
rectified linear units(ReLU)
另一种激活函数(activation function),f(z)≡max(0,z)
比sigmoid性能更好
理由:在z较大时不会发生饱和,但目前尚没有比较详细的解释。
扩展训练数据集
再插入一个全连接层,并加入dropout正则化(dropout仅用于全连接层,不必用于卷积层,因为卷积层的共享权重本身就带有一些正则化)
神经网络的集成(ensemble)
hysic
commented
7 years ago
图像识别(image recognition)经典论文
2012 LRMD paper
, ImageNet -> 15.8%
2012 KSH paper
, aka AlexNet, ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) -> top-5 84.7%, top-1 63.3%
2014 ILSVRC competition
, aka GoogLeNet
其他图像识别项目
ImageNet
google用深度CNN识别谷歌街景中的门牌号
为正确分类的图片叠加一个adversarial image,图片将无法正确分类,而肉眼却几乎看不出差别
与之相对,一些肉眼看起来完全是噪声的图片,神经网络却可以以很好的置信度将其划归某个标签
hysic
commented
7 years ago
其他深度神经网络
递归神经网络(Recurrent neural networks, RNNs)
长短期记忆网络(Long short-term memory, LSTM)
深度置信网络(Deep belief nets, DBN),生成模型(generative models),玻尔兹曼机(Boltzmann machines)
hysic
commented
7 years ago
神经网络的未来
Intention-driven user interfaces:例如谷歌搜索的自动纠错,siri
Machine learning, data science, and the virtuous circle of innovation: find the "known unknowns" hidden in data
卷积神经网络(CNN)