- Chestnut: Predict the price of the house based on the size of the house.
(figure 1)
- Description: The straight line fits the sample points (linear regression) but the price is not negative, so the front is a line that tends to zero.
- The above example can be understood as a simple neural network:
(Figure 2) (Single Neuron Network)
size of house (input x) - (neuron node) - price (output y)
- The fitted curve function on the right side of Figure 1 is called: ReLU function (rectified linear unit) Correction means to deal with the case where the price is not negative
- The neural network can be regarded as a collection of multiple single-neuron networks. There are many factors that affect the output (for example, the factors that affect the price of houses are not only the size of the house, but also the number of people living in the buyer's family, etc.), the input layer can be There are multiple features x, but these features need to be mapped to the output y through neurons.
(image 3)
1.3 Supervised Learning with Neural Networks
- Supervised learning: A type of machine learning.
Application Scenario: Precise placement of advertisements on users / Unmanned driving (judging whether there is a car around) / Translation from English to Chinese /...
(Figure 4)
Choose different neural network systems for different scenarios:
Housing price prediction/ad placement: Standard neural network architecture Standard NN
Image: Convolutional Neural Network CNN
Audio: Recurrent Neural Network (Recurrent Neural Network) (the time component in audio can be regarded as sequence data/one-dimensional time series)
Text: RNNs (letters, etc. appear one by one, which can be regarded as sequence data)
(Figure 5 SNN)
(Figure 6 CNN)
(Figure 7 RNN)
- Machine learning is used for both structured and unstructured data - structured data is data from a database; unstructured data (audio/image/text etc.)
1.4 Why is deep learning emerging?
(Figure 8)
- More and more data provides the basis for deep learning analysis and training samples
- Rapid development of network conditions/CPU and other hardware
- Algorithms that continue to innovate