Скачать или смотреть [MXDL-13-01] Autoencoder [1/6] - Dimensionality reduction

In this series, we will look at autoencoders. In this video, we will look at the basics of autoencoders and dimensionality reduction using autoencoder models.

Let's look at the full table of contents for this series.

In Chapter 1, we will look at the basic concepts of autoencoders.

In Chapter 2, we will look at dimensionality reduction using autoencoders. We will build fully connected autoencoder and CNN autoencoder models to reduce the dimensionality of MNIST images.

In Chapter 3, we will look at noise reduction using autoencoders. We will build a CNN autoencoder model to reduce noise and motion blur in MNIST images.

In Chapter 4, we will look at sparse autoencoders. Let's impose sparsity constraints on autoencoders by applying KL divergence or L1 activity regularization. And let's apply a sparse autoencoder to the MNIST dataset.

In Chapter 5, we will look at anomaly detection. Let's use an autoencoder model to perform credit card fraud detection.

Finally, in Chapter 6, we will look at variational autoencoder (VAE), a type of generative model. Let's generate MNIST images using a variational autoencoder model.

An autoencoder is a network trained so that its output is as similar as possible to the input data. An autoencoder neural network is an unsupervised learning algorithm that applies backpropagation. Autoencoders have two main components: an encoder and a decoder. The encoder typically reduces the dimensionality of the input data x to produce a latent vector z. Conversely, the decoder reconstructs the original data x from the latent vector z, which is a reduced representation of x.

The goal of an autoencoder is to minimize the difference between the original input and the reconstructed output, a measure known as the reconstruction error. The reconstruction error can be simply measured by the mean squared error.

An autoencoder can be configured as a single layer, or as multiple layers, called a stacked autoencoder. Stacked autoencoders allow the network to learn hierarchical representations of the input data x, with each layer capturing more complex features.

The concept of autoencoder was first introduced in the 1980s, early works on autoencoder were used for dimensionality reduction or feature learning. Since then, many types of autoencoder have been proposed by different researchers and successfully applied in many fields, such as generative models, computer vision, speech recognition and natural language processing. Autoencoders can be applied to dimensionality reduction, noise removal, anomaly detection, generative models, etc.

#Autoencoder #DimensionalityReduction #Conv2DTranspose