Deep Learning
Copyright By PowCoder代写 加微信 powcoder
basic concept of
machine learning
representation learning
find good feature space automatically instead of hand-crafted feature
to achieve high performace with a small dataset
some case will not happen
umanifold assumption apply smooth change according to certain rules
underfitting
model capacity is too small to fit the data accordingly
use model with higher order
over-fitteing
trainning process also accepts data noise
low generalization ability
type of supervision
supervised learning both feature and the output are given
unsupervised learning the feature is given, but the output is not given
semi-supervised learning
some data have both feature and outpu but others only have feature
in some cases, it’s easy to get feature but outputs hould requires manual tasks
reinforcement learning
agent perform the action and if the action is correct, it will give reward.
e.g. Alpha Go
Deep Learning
end-to-end traning
learning features/representations
learning multi-level features and an output different from representation learning
neural network based model
a series of linear classifiers, non-linear activations and loss function
cascadeing the neurons to form a neural network
Loss function
Hinge loss
cross-entropy loss
log likelihood loss
regression loss
sj for score of the false label
syi for score of ture label
L=sum(Li)/N 求平均
for classification
for classification
for contiuous output
or yi for label and si for prediction
Active Function
对每个output计算exp(output)，并归一化
Regularization avoid overfitting
gradient descent
backpropagation
Receptive field the region of the input space that affects a particular unit of the network
convolutional filter size
three 3*3 Conv layers and single 7*7 Conv layer have the same receptive
but the small filter size produce more expressive activation map, which is better
large filter size has more parameters
spatial dimention (size of input – size of kernel + 2*padding)/stride + 1
1*1 Conv can reduce depth size
case studies
small fitler, Deeper network
Inception module
Naive Inception module Total depth after concatenation can onlhy grow at every layer
Improved Inception module bottleneck layers which use 1*1 convolutions to reduce feature depth
Residual block
difficult for optimization
AlexNet use one 7*7 Conv layer
VGG use three 3*3 Conv layer
the first CNN-based winner
batch size the number of input images/data
倒数是(1-s)*s
Convolution layer input 100*100*3, 32 kernels 3*3*3 reshape to 27*10000 and 32*27
regularization/use large dataset for training/Dropout
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com