LiSSA: Localized Stochastic Sensitive Autoencoders.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Autoencoder (AE) training aims to minimize training error and regularization terms for generalization.
• Current AE training methods may lack robustness to input perturbations, hindering generalization.
• Poor generalization capability is a significant challenge in autoencoder applications.

Purpose of the Study:

• To propose a novel autoencoder method, localized stochastic sensitive AE (LiSSA), to enhance robustness against input perturbations.
• To improve the generalization capability of autoencoders for unseen data.
• To develop a more robust feature learning method for classification tasks.

Main Methods:

• Introduced localized stochastic sensitivity regularization in autoencoder training.
• Developed LiSSA to reduce sensitivity to small input perturbations.
• Preserved local connectivity from input to representation space for robust feature learning.

Main Results:

• LiSSA demonstrated enhanced robustness to input perturbations.
• Learned features from LiSSA improved classifier generalization capability.
• LiSSA significantly outperformed classical and recent AE training methods on 36 benchmarking datasets.

Conclusions:

• LiSSA effectively enhances the robustness and generalization of autoencoders.
• The proposed method offers a significant improvement over existing AE training techniques.
• LiSSA provides a promising approach for robust feature learning in classification.