Disentanglement Framework
Disentanglement frameworks aim to decompose complex data representations into independent, interpretable latent factors, improving model understanding and robustness. Current research focuses on developing novel architectures, such as variational autoencoders and contrastive learning methods, often incorporating adversarial training or mutual information minimization to achieve effective disentanglement across diverse data modalities (e.g., speech, images, time series). This work is significant for enhancing the interpretability of deep learning models and improving the performance and reliability of applications ranging from speaker recognition and pathology analysis to face forgery detection and keyword spotting.
Papers
Disentangled Training with Adversarial Examples For Robust Small-footprint Keyword Spotting
Zhenyu Wang, Li Wan, Biqiao Zhang, Yiteng Huang, Shang-Wen Li, Ming Sun, Xin Lei, Zhaojun Yang
Toward Improving Synthetic Audio Spoofing Detection Robustness via Meta-Learning and Disentangled Training With Adversarial Examples
Zhenyu Wang, John H. L. Hansen