深度学习：
1 three steps for deeping learning:
define a setof function–>goodness of function–>pick the best function
1-1 neural network
parameters
each neurons can have different values of weights and biases.
feed forward
vector
1-2 three layers:
input layer–>hidden layers–>output layer
trial+error+intuition
find a function in function set that minimizes total loss.
1-3 gradient descent
repeat
backpropagation
2 overfitting:
when using softmax output layer,choose cross entropy.
mini-batch
epoch:时间上的一个点
parallel computing
shuffle:洗牌
3 derivative:衍生物
versa:反的
momentum:动量
stuck:被卡住的
saddle:鞍状物
plateau:高原
prune:砍掉
synaptic density:突触密度
decay:衰退
implementation:实现
variant:不同的
filter:滤过
matrix:矩形
detect:探测
stride:步幅
supervised learning:监督式学习
reinforcement learning:加强学习
unsupervised learning:无监督式学习
ensemble:同时
semantic:语义的
scenario:方案
sacrifice:牺牲
retrieval:检索
4 property:
some patterns are much smaller than the whole image.
the same patterns appear in different regions.
subsampling the pixels will not change the object.