Class Aware Autoencoders for Better Feature Extraction

This conference paper presents a novel approach to improving feature extraction in autoencoders by integrating class labels into the training process. Unlike traditional autoencoders that operate unsupervised, this method enhances the autoencoder’s ability to learn features that are representative of both individual data points and their respective classes.

The effectiveness of this approach is validated by comparing the accuracy of classifiers trained on features extracted by the proposed class-aware autoencoders against those trained on features from traditional autoencoders. Experiments conducted on the MNIST, CIFAR-10, and UTKFace datasets demonstrate that the classifiers achieve higher accuracy when utilizing features extracted by the class-aware autoencoders. This research highlights the potential of incorporating class information into the feature extraction process to improve model performance.

Recommended citation: Islam, A., & Belhaouari, S. B. (2021). "Class Aware Autoencoders for Better Feature Extraction." In 2021 International Conference on Electrical, Communication, and Computer Engineering (ICECCE) (pp. 1--5). IEEE. https://doi.org/10.1109/ICECCE52056.2021.9514202
Download Paper