Autoencoders: Deep Dive
Last updated
Last updated
Most of knowledge derived from this source.
Variational Autoencoders explained in-depth here.
Here, we'll tackle how the autoencoder works and how it compares to traditional linear approaches such as principal component analysis (PCA) for computer vision and latent semantic analysis (LSA) for natural language processing.
An autoencoder is an unsupervised machine learning technique that utilizes a neural network to produce a low dimensional representation of a high dimensional input.
PCA finds the directions of maximal variance in high-dimensional data. By selecting only those axes that have the largest variance, PCA aims to capture the directions that contain the most information about the inputs, so we can express as much as possible with a minimal number of dimensions. The linearity of PCA, however, places significant limitations on the kinds of feature dimensions that can be extracted. Autoencoders overcome these limitations by exploiting the inherent nonlinearity of neural networks.
An autoencoder consists of two major parts, the encoder and the decoder networks. The encoder network is used during both training and deployment, while the decoder network is only used during training. The purpose of the encoder network is to discover a compressed representation of the given input.The purpose of the decoder network, which is just a reflection of the encoder network, is to reconstruct the original input as closely as possible. It is used during training in order to force the autoencoder to choose the most informative features in its compressed representation. The closer the reconstructed input is to the original, the better our original representation!
If we were trying to build a generative model using the method above, we can't generate anything yet, since we don't know how to create latent vectors other than encoding them from images.
There's a simple solution here. We add a constraint on the encoding network, that forces it to generate latent vectors that roughly follow a unit gaussian distribution. It is this constraint that separates a variational autoencoder from a standard one.
There's a tradeoff between how accurate our network can be and how close its latent variables can match the unit gaussian distribution.
We let the network decide this itself. For our loss term, we sum up two separate losses: the generative loss, which is a mean squared error that measures how accurately the network reconstructed the images, and a latent loss, which is the KL divergence that measures how closely the latent variables match a unit gaussian.
In order to optimize the KL divergence, we need to apply a simple reparameterization trick: instead of the encoder generating a vector of real values, it will generate a vector of means and a vector of standard deviations.
Here's something convenient about VAEs. Since they follow an encoding-decoding scheme, we can compare generated images directly to the originals, which is not possible when using a GAN.
A downside to the VAE is that it uses direct mean squared error instead of an adversarial network, so the network tends to produce more blurry images.
There's been some work looking into combining the VAE and the GAN: Using the same encoder-decoder setup, but using an adversarial network as a metric for training the decoder. Check out this paper or this blog post for more on that.
You can follow a really nice code walk-through here!
