@incollection{gcvae_talk,
  doi = {10.48550/ARXIV.2206.04225},
  url = {https://arxiv.org/abs/2206.04225},
  author = {Ezukwoke, Kenneth and Hoayek, Anis and Batton-Hubert, Mireille and Boucher, Xavier},
  keywords = {invited},
  title = {Artificial intelligence for Fault Analysis in Semiconductor Industry. Highlight: Generalized-Controllable Variational AutoEncoder (GCVAE)},
  publisher = {Plate-Forme Intelligence Artificielle (PFIA-22)},
  year = {2022}
}

Variational autoencoders (VAEs) have recently been used for unsupervised disentanglement learning of complex density distributions. Numerous variants exist to encourage disentanglement in latent space while improving reconstruction. However, none have simultaneously managed the trade-off between attaining extremely low reconstruction error and a high disentanglement score. We present a generalized framework to handle this challenge under constrained optimization and demonstrate that it outperforms state-of-the-art existing models as regards disentanglement while balancing reconstruction. We introduce three controllable Lagrangian hyperparameters to control reconstruction loss, KL divergence loss and correlation measure. We prove that maximizing information in the reconstruction network is equivalent to information maximization during amortized inference under reasonable assumptions and constraint relaxation.

@incollection{gcvae_tall,
  doi = {10.48550/ARXIV.2206.04225},
  url = {https://arxiv.org/abs/2206.04225},
  author = {Ezukwoke, Kenneth and Hoayek, Anis and Batton-Hubert, Mireille and Boucher, Xavier},
  keywords = {regular},
  title = {Deep Generative Model. Highlight: Generalized-Controllable Variational AutoEncoder (GCVAE)},
  publisher = {Seminar at GMI/DSI-LIMOS/MAAD-LIMOS, Mines Saint-Étienne},
  year = {2022}
}

Variational AutoEncoders (VAEs) have recently been used for unsupervised disentanglement learning of overlapping mixtures of distributions (e.g Gaussian). Numerous variants exist to encourage disentanglement in hidden (latent) spaces while improving reconstruction quality. However, none have simultaneously managed the trade-off between attaining extremely low reconstruction error and a high disentanglement score. We present a generalized framework to handle this challenge under constrained optimization and demonstrate that it outperforms state-of-the-art existing models as regards disentanglement while balancing reconstruction. We will begins first by introducing you to the formalization of variational autoencoders and their closed form solution followed by a more stable version, Generalized Variational Autoencoder (GCVAE).