Easyai notes - deep learning

CNN

CNN The value of ：

1. To be able to ⼤ Graph of data volume ⽚ Effective dimensionality reduction ⼩ Data volume ( Does not affect results )
2. Be able to keep the graph ⽚ Characteristics of , similar ⼈ The visual principle of class

CNN The basic principle of ：

1. Convolution layer – The main work is ⽤ It's a reserved graph ⽚ Characteristics of
2. Pooling layer – The main work is ⽤ Is to reduce the dimension of data , Can effectively avoid over fitting
3. Fully connected layer – Output the results we want according to different tasks

CNN The actual should ⽤：

1. chart ⽚ classification 、 retrieval
2. ⽬ Mark location detection
3. ⽬ Mark segmentation
4. ⼈ Face recognition

RNN

RNN With traditional nerves ⽹ Collaterals ⼤ The difference between them is that they always put the former ⼀ The output of times , Take it to the next ⼀ Secondary hidden layer .

⻓ Short term memory ⽹ Collateral – LSTM
In standard RNN in , Only a single tanh layer , Keep only important information .

GRU yes LSTM Of ⼀ Variants . Retain the LSTM Focus on , The characteristic of forgetting unimportant information , stay long-term It's not lost when it's spread . In the training dataset ⽐ a ⼤ Can save a lot of time .

RNN Its unique value lies in ： It can effectively process sequence data .
be based on RNN There is LSTM and GRU And so on . These variant algorithms mainly include ⼏ Characteristics ：

1. ⻓ Period information can be effectively retained
2. Select important information to keep , Unimportant information will choose “ Forget ”
   RNN ⼏ A typical answer ⽤ as follows ：
3. ⽂ Ben ⽣ become
4. language ⾳ distinguish
5. Machine translation
6. ⽣ Image description
7. Video Tags

GANs

The original ：⾃ Dynamic
⽣ Become a confrontation ⽹ Collateral （GANs） from 2 An important part of ：

1. ⽣ Make it (Generator)： By machine ⽣ It's data （⼤ In some cases, images ）,⽬ Yes. “ Cheated ” Judging device
2. Judging device (Discriminator)： Judge whether this image is real or machine ⽣ Yes ,⽬ The key is to find ⽣ become
   It's made of “ Fake data ”

3 Advantages

1. Can better model data distribution （ The image is sharper 、 Clear ）
2. Theoretically ,GANs Can train any ⼀ Kind of ⽣ Make it ⽹ Collateral . Other frameworks need ⽣ Make it ⽹ Collaterals ⼀ Some special
   Definite functional form ,⽐ If the output layer is ⾼ Si .
3. ⽆ Need interest ⽤⻢ Markov chain repeated sampling ,⽆ Need to advance in the learning process ⾏ infer , Without complex variations
   world , Avoid approximate calculation thorns ⼿ The problem of probability .

2 A flaw

1. Hard to train , unstable .⽣ A good synchronization is needed between the generator and the discriminator , But in practice it's easy
   D convergence ,G Divergence .D/G Your training needs to be refined ⼼ The design of the .
2. Pattern missing （Mode Collapse） problem .GANs There may be a lack of patterns in the learning process ,⽣ Make it
   Begin to degenerate , Always ⽣ Into the same sample point ,⽆ Can't continue learning .