Deep learning notes

---

1. Skills needed

1. Programming skills -python

about python,　 In addition to mastering its own grammar , You should also focus on the following libraries ：
pandas： Super excel, Tabular operation data , Powerful tools for data cleaning and preprocessing .
numpy： Numerical calculation library , Quick, don't, don't .
matplotlib： imitation MATLAB Data visualization tools for .
scikit-learn： Encapsulate super good machine learning library , Some simple algorithms are not easy to use .
ipython notebook： Notebooks for data scientists and Algorithm Engineers , Strongly recommend .

2. Mathematical basis

Know partial differential and its properties 、 The concept of gradient , That is, Tongji high number one + The first few chapters of high number two are linear algebra —— Matrix four / Differential operation , Linear space 、 linear transformation 、 Eigenvalue eigenvector probability theory —— probability 、 Conditional probability 、 Bayes' formula 、 Probability distribution optimization problem —— Mainly gradient descent method

3. Image processing part

Image processing the basic method of image processing —— wave filtering 、 Two dimensional convolution 、 Two dimensional orthogonal transformation 、 Basic morphological operations （ open / Closed operation , inflation / corrosion ）

4. Convolutional neural networks (CNN)

Convolutional neural networks are divided into Convolution layer 、 The nonlinear layer 、 Pooling layer 、 Fully connected layer .

Fully connected layer Observe the output of the previous layer （ It represents the activation mapping of higher-level features ） And determine which classification these features are most consistent with .

The computer adjusts the filter value through a training process called back propagation （ Or weight ）.

Back propagation It can be divided into four parts , Namely Forward feedback 、 Loss function 、 Backward feedback , as well as Weight update .

Forward feedback ： adopt Convolution operation 、 Merge operation 、 Activation function ( Nonlinear mapping ) perhaps Other operating / Transformation Wait for a series of operations to stack layer by layer , Extract the high-level semantic information from the original data input layer layer by layer , Pay attention to abstraction .

Loss function ： Calculation error .

Backward feedback ： Determine which part of the weight caused the greatest loss , Find ways to adjust and reduce losses .

Weight update ： Once the backward derivative is calculated , The last step is to update the weight . The weights of all filters will be updated , So that they change along the gradient .

Convolution What is taken is Local features ;
Full connection Namely Reassemble the previous local features into a complete graph through the weight matrix .
Pooling layer （ Just maximum pooling ） It's like “ A ticket ” The role of , Different characteristics in different “ The candidate ” Have their own preferences .

---

2. CNN Related information

You know ： CNN（ Convolutional neural networks ） What is it? ？ Do you have any introduction or articles ?
You know ： YJango Convolution neural network of —— Introduce

---

3. object detection

YOLO Direct adoption regression（ Return to ） Methods Detect and classify coordinate frames , Use one end-to-end Simple network , Directly realize coordinate regression and classification , It means trained YOLO When you input an image , It can directly output the coordinate frame containing the object and what the object is .

YOLO V1 Information ： < Machines love learning >YOLO v1 In depth understanding of

YOLO Brief introduction of the framework