High-dimensional and high-complexity unsupervised abnormal data recognition with neural networks remains an open research question. We propose a multi-level Variational Auto Encoder (VAE) structure in this work. First, based on traditional VAE, we deconstruct the original data using convolution kernels of different specifications to establish multiple feature latent spaces, thereby reducing the excessive data aggregation in the latent space. Then we use an equal-window convolution kernel to fuse feature blocks of different sizes, thereby transforming local features into globally identifiable features. Moreover, the model structure shortens the distance between the sub-feature block and the output end-point, reducing the training inefficiency caused by the vanishing gradient.
Multi-level Variational Auto Encoder for Unsupervised Abnormal Data Recognition
27.06.2022
1246690 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Disentangling Latent Factors of Variational Auto-Encoder with Whitening
ArXiv | 2018
|Causal Intervention Framework for Variational Auto Encoder Mechanistic Interpretability
ArXiv | 2025
|