Deep learning (Deep Learning) is Machine Learning (Machine Learning) in a new field of research, led the third wave of artificial intelligence.
This paper analyzed three typical areas of deep learning algorithms, hope to help you better understand the depth of learning this very deep subject.
1. convolutional neural network (CNN)
Convolutional neural network (Convolutional Neural Networks, CNN) is a class that contains convolution calculation possessed deep structure of feed-forward neural network (Feedforward Neural Networks), it is one of the representatives of deep learning algorithm.
Convolution neural networks mimic biological visual perception (Visual Perception) mechanism to build, can be supervised learning and unsupervised learning.
CNN typically consists of three parts - the convolution layer, pooled layer, fully connected layer.
Wherein the convolutional layer is responsible for extracting local feature in the image; pooling layer serves to significantly reduce the order parameters (dimension reduction); layer fully connected neural network similar to a conventional section for outputting a desired result.
CNN in image processing is very advantageous, currently retrieving the image classification, targeting detection, object segmentation, face recognition, recognition bones and other fields have a wide range of applications.
2. Recurrent Neural Network (RNN)
Recurrent Neural Network (Recurrent Neural Network, RNN) is a type recurrent neural network, and all nodes recursively (circulation means) in the direction of the evolution of sequence data entered in the sequence of chained.
Recurrent Neural Networks has memory, and the parameter sharing Turing complete (Turing Completeness), therefore has certain advantages when the nonlinear characteristics of the learning sequence.
Depth field study, RNN is an effective algorithm for sequence data processing. In text generation, speech recognition, machine translation, generating a field of image is described, the video tag have a wide range of applications.
3. Generate confrontation Network (GAN)
Generated against the network (GAN, Generative Adversarial Networks) is a deep learning model is very popular the last two years of an unsupervised learning algorithm.
Generated against the network (GAN) consists of two important parts:
1. Generator (Generator): generates data (image in most cases) through the machine, the purpose of the "fool" the discriminator;
2. discriminator (Discriminator): judging this image is real or machine generated, the purpose is to identify the builder to do "false data."
GAN has the following three advantages:
1. better distribution modeling data (image sharper, clearer);
2. Theoretically, GAN can train any network generator;
3. repeated without using the Markov chain sample, and need not be estimated in the learning process.
But there are two drawbacks:
1. hard training, you need a good synchronization between instability, the generator and the discriminator.
2. Lack mode. GAN mode of learning may appear missing, the generator starts to degrade, always generate the same sample point, unable to continue their studies.
GAN can generate a very realistic photos, images and even video, generated image data sets to produce human face photograph, image to image conversion, convert text to images, image editing, picture restoration, and many other fields have a wide range of applications.