As mentioned earlier, a regularized autoencoder extends the standard autoencoder by adding a regularization parameter to the cost function, shown as follows:
Here, λ is the regularization parameter and i and j are the node indexes with W representing the hidden layer weights for the autoencoder. The regularization autoencoder aims to ensure more robust encoding and prefers a low weight h function. The concept is further utilized to develop a contractive autoencoder, which utilizes the Frobenius norm of the Jacobian matrix on input, represented as follows:
where J(x) is the Jacobian matrix and is evaluated as follows:
For a linear encoder, a contractive encoder and regularized encoder converge to L2 weight decay. The regularization helps in making the autoencoder less sensitive to the input; however, the minimization of the cost function helps the model to capture the variation and remain sensitive to manifolds of high density. These autoencoders are also referred to as contractive autoencoders.