First week: deep learning Introduction
1.1 Introduction The depth learning
to change the traditional Internet business: Internet search advertising.
Better to do: reading x-rays, individualized education, precision agriculture
cousera course
second course: Super parameters, regularization, diagnosis deviation, variance, advanced optimization algorithms.
Advanced optimization algorithms: momentum, adam algorithm
third course: Construction Strategy machine learning systems vary the depth learning error, end-to-depth study
of the fourth course: convolution neural network (CNNS), apply image.
Fifth course: series model, application of natural language processing.
Series model: Recurrent Neural Network (RNN), short and long term memory network (LSTM)
1.2 Neural Network
ReLU activation function: rectified linear unit. rectify correction max (0, x)
1.3 Supervised learning neural networks
present all the economic value created by the neural network, the essence of supervised learning.
application:
- ----- online advertising website advertising input information, and user information, to consider whether the site to show ads.
- Predicting whether the user opening the ad, the user may recommend to click on ads
- Computer Vision
- Speech recognition: the voice input, text output
- Machine Translation: input English sentence, the output of Chinese sentences
- Autopilot: input pictures, find out what the car in front. Tell specific car on the road position
Image: convolution (CNN)
Sequence data: Recurrent Neural Network (RNN), more complex (RNNs). Data are part of the time. (Such as audio, there is a time component over time, audio playback.)
Application details:
- Structured data: a basic database
- Non-data of the data: hard, speech recognition, image recognition, natural language processing
Short-term economic value creation: structured data
1.4 depth study the rise of
accuracy: spam filtering, click advertising forecast, autonomous vehicles determine the location of
the neural network a huge breakthrough: the sigmoid function -> ReLU function
sigmoid function: gradient descent and gradient close to zero, the parameter update slow, slow learning speed.
ReLU function: linear correction means. Negative gradient of 0 input, the gradient does not tend to decrease gradually 0