Ya-lin/DAE-vs-CAE

Autoencoders (AEs)

An autoencoder learns two functions: a mapping $g$ from $\mathbb{R}^D$ to $\mathbb{R}^d$, called the encoder, and another mapping $h$ from $\mathbb{R}^d$ to $\mathbb{R}^D$, called the decoder. The model is trained to minimize

$$\mathcal{L}=\|h(g(\mathbf{x}))-\mathbf{x}\|_2^2.$$

More generally, we can use

$$\mathcal{L}=-\ln p(\mathbf{x}|h(g(\mathbf{x}))).$$

Applying AE in any application, we have to restrict the model in some way such that it does not learn the identity function.

Bottleneck autoencoders

The simplest approach to restrict the mode is using a narrow bottleneck layer, i.e., $d< D$. The bottleneck autoencoders extract salient features, typically for dimensionality reduction. Autoencoders with nonlinear encoder functions $g$ and nonlinear decoder functions $h$ can thus learn a more powerful nonlinear generalization of PCA. Unfortunatly, if the encoder and decoder are allowed too much capacity, the autoencoder can learn to perform copying task without extracting useful information about the distribution of the data. Theoretically, one could imagine that an autoencoder with a one-dimensional code but a very powerful nonlinear encoder could learn to represent each training example $\mathbf{x}_i$ with the code $i$. The decoder could learn to map these integer indices back to the values of specific training examples. These specific scenario does not occur in practice, but it illustrates clearly that a bottleneck trained to perform the copying task can fail to learn anything useful about the dataset if the capacity of the autoencoder is allowed to become too great.

Denoising autoencoders (DAEs)

A denoising autoencoder is a modification of the autoencoder to prevent the network from learning the identity function. Specifically, if the autoencoder is too big, then it can just learn the data, so the output equals the input, and does not perform any useful representation learning or dimensionality reduction. Denoising autoencoders solves this problem by corrupting the input data on purpose, adding noise, or masking some of the input values. The autoencoders are trained to reconstruct the original images (noise-free images) by taking corrupted or masked images. Let $g$ and $h$ be the encoder and decoder network. Then the loss function in DAE is given by

$$\min_{g, h}\frac{1}{n}\sum_{i=1}^n\left\|h(g(\mathbf{x}_i^{\epsilon}))-\mathbf{x}_i\right\|_2^2$$

where $\mathbf{x}_i^{\epsilon}$ is the corrupted version of $\mathbf{x}_i$. For image dataset, usually $\mathbf{x}_i^{\epsilon}$ is obtained by adding independent Gaussian noise and truncating values into $[0,1]$ (assume pixel values are normalized), i.e.,

$$\mathbf{x}_{i}^{\epsilon}=\textbf{clip}_{to [0,1]}\left(\mathbf{x}_i+\epsilon_i\right).$$

where $\epsilon_i\sim\mathcal{N}(0,\sigma^2)$.

Remark.

• DAE aims to remove noise from corrupted images. So it could be an overcomplete autoencoder, i.e., $d>D$.
• Don't forget to clip images after adding noise.
• Masking images is an alternative to corrupting images by Gaussian noise.

Contractive autoencoders (CAEs)

A different way to regularize autoencoders is by adding the penalty term

$$\Omega(\mathbf{z},\mathbf{x})=\lambda\left\|\frac{\partial g(\mathbf{x})}{\partial\mathbf{x}}\right\|_F^2=\lambda\sum_k\|\nabla _{\mathbf{x}}g_k(\mathbf{x})\|_2^2$$

to the reconstruction loss, where $g_{k}(\mathbf{x})$ is the value of $k'$th component of $g(\mathbf{x})$. That is, we penalize the Frobenius norm of the encoder's Jacobian. Remark. To minimize the penalty term, the model would like to ensure the encoder is a constant function. However, if it was completely constant, it would ignore its input, and hence incur high reconstruction loss. Thus the two terms together encourage the model to learn a representation where only a few units change in response to the most significant variations in the input. Unfortunately CAEs are slow to train, because of the expense of computing the Jacobian.