Deep learning, as the name implies, is a deeper level of learning. It is a subfield of machine learning. As far as I understand it, it is a deep neural network.
So what is a deep neural network? As the name implies, it is a neural network with many hidden layers. Usually, as long as there are more than 2 hidden layers, we can use this neural network
The network is defined as a deep neural network. Of course, the activation function of the hidden layer should be a nonlinear function. If it is linear, even if it has 10,000 layers, its learning ability
It is also just equivalent to a neural network with one hidden layer.
What are the benefits of deep learning, or what advantages does it have in terms of traditional machine learning? Personally, I think its greatest benefit is that it can learn from raw data.
Features are automatically extracted from the data without the need to manually design features. This is also called feature learning or representation learning, and in representation learning, the most important thing is to
Layer processing, then what is layer-by-layer processing? If we want to process an image, the neural network has many layers, then at the bottom layer, the network sees pixels. When
When going up layer by layer, what you see may be an edge, and if you go up, it may be an outline or a part of a certain target. What the neural network does is to look at it layer by layer.
To abstract things, combine low-level features into high-level features. When it comes to deep neural networks, we have to talk about wide neural networks, as the name suggests, wide neural networks
The network is to widen a hidden layer, that is, to increase the number of neurons in the hidden layer. Although it also increases the complexity of the model, the effect is not as effective as a deep neural network.
Well, one of the main reasons is that it is not processed layer by layer. Compared with other traditional machine learning, deep learning also provides complex feature transformation, traditional machine learning
For example, decision trees, although they can also reach very deep, are always learned in a feature space, and no feature transformation is performed internally, in other words, they are processing data
Only raw features are used, and deep neural networks can perform complex feature transformations on these raw features to learn more advanced features.
Deep learning is not a panacea. Many people may know that deep learning has a strong representation ability, so it can be used to solve any problem. This kind of thinking is very dangerous. generally speaking
Deep learning is more suitable for processing structured data, traditional machine learning is more suitable for processing unstructured data, structured data is images, videos, natural language, etc. They have a
The commonality is that these data are composed of smaller granularities. For example, images are composed of pixels, sounds are composed of phonemes, and sentences in text are composed of words.
Low-level features are combined into high-level features, and local features have strong correlations. Deep learning can process layer by layer, so it is very suitable for processing these structured data rather than structured data.
Data, such as booking a hotel, buying a house, etc., will show the characteristics of those houses or hotels, such as how many bedrooms, how many bathrooms, and the height of the ceiling. It is obvious that
These features can no longer be combined into high-level features, and there is no strong local correlation between features, so traditional machine learning is suitable for dealing with these problems.
The original feature space for learning. But it is not that the deeper the network, the stronger its learning ability, and some other problems will be mentioned later.