Machine Learning 001 Deeplearning.ai Depth Learning course neural Networks and deep learning first week summary

Deep Learning Specialization

Wunda recently launched a series of courses on deep learning in Coursera with Deeplearning.ai, which is more practical compared to the previous machine learning course. The operating language also has MATLAB changed to Python to be more fit to the current trend. A study note on this series of courses will be made here.

The deep learning specialization is divided into five courses, namely: Neural Networks and deep learning,improving deep neural Networks: Hyperparameter tuning, regularization and optimization,structuring machine learning projects,convolutional Neural Networks,sequence Models.

This essay begins with the first week of neural Networks and deep learning, first course

Introduction to Deep learning

Learning Goals:

1. Understand the key trends driving deep learning

2. To explain how deep learning is applied to supervised learning.

3. Learn about the major categories of deep learning models and when to apply what models

4. Easy to understand what happens when machine learning is working well

Main content

Artificial intelligence will bring about another great change to human society like the advent of electricity. Just like 100 years ago, AI is also changing many industries.

What is a neural network

Suppose we use the size of the house to predict the price of a house, then we can use linear regression (linear regression) to fit, and we know that the price of the House will not be a negative value, so we can use the Blue line to do a fitting.

This blue line is very similar to the current active function Relu (rectified linear unit) commonly used in neural networks. For this problem with only one input and one output, we can use one of the simplest neural networks to represent.

This simplest neural network is like Lego bricks, and more complex networks can be stitched together with this simplest "Lego" block.

Of course, we may face more complex problems in our daily lives. The factors that determine the price of a house include, in addition to the size of the house, the number of bedrooms, the postal code (the street where the house is located) and the affluence of the surrounding communities. The size of the house and the number of bedrooms can determine if the house is suitable for a few homes, and the streets of the House may determine whether the traffic is convenient, the streets of the houses, and the affluence of the surrounding communities may decide whether the house is a school district room or not. Then the house is suitable for a few homes, the traffic is convenient, whether it is the school district room to predict the price of the house. When we use neural networks to solve such problems, we only need to give input x that is, the size of the house, the number of bedrooms, zip code and the surrounding affluence can be obtained through a trained neural network to get a forecast house price.

From this neural network we can see that each feature of the input layer is connected to each neuron of the hidden layer. Of course, the weight of the connection may be different.

Supervised learning using a neural network

Supervised learning, we are bound to have the right input x and output Y. and supervised learning has a wide range of applications at present. In different applications, we should choose the appropriate input and output

In different fields of application, we may take different deep learning models. In real estate (real estate) and online advertising, for example, we prefer to use standard neural networks (STANDARDNN). We often use convolutional neural networks (CNN, convolutional neural network) to mark images or image recognition. In the sequence recognition of speech recognition and translation, the commonly used models are recurrent neural networks (RNN, recurrent neural network). When faced with complex problems such as autonomous driving, we may use complex hybrid networks (complex hybrid neural network architecture.).

In supervised learning, we may face two kinds of data: structured data and unstructured data. For structured data, there are clearly defined meanings for each feature. Unstructured data, such as sound, images, and text, can be characterized by a single syllable or pixel, with no definite meaning.

What drives deep learning to soar

1. Deep learning has achieved great success in many applications, such as online advertising recommendations, speech recognition, and image recognition.
2. At present, the computing power of computers has improved greatly compared with the past.
3. We now have a lot of data (big data)

From this graph we can see that more data, larger networks can improve the performance of neural networks. At the same time, the development of the algorithm cannot be neglected. For example, from the sigmoid activation function used in the past to the current common Relu activation function. At the same time, the new algorithm generally makes the neural network training speed faster. Faster training speeds give us the opportunity to train larger networks to cope with more and more data, while also allowing us to try more ideas to test different parameters faster. The development of hardware also allows us to train the network model more quickly to validate our ideas.

Machine Learning 001 Deeplearning.ai Depth Learning course neural Networks and deep learning first week summary