Matrix-Variate Beta Variational Autoencoder

Variational autoencoders (VAEs) are currently popular deep generative models that demonstrate powerful performance in various applications. However, VAEs have a tendency to vectorize data during encoding, which can result in the loss of structured information for matrix-based data, such as images. Some researchers have extended VAEs to develop matrix-based variational autoencoder methods, which have shown improved performance in handling image data. Building upon this work, we propose Matrix-variate beta VAEs, which can extract latent representations from raw matrix data in a fully unsupervised manner and effectively preserve the spatial information of the data. Through experiments conducted on standard image datasets, our method outperforms traditional vectorized and matrix-based VAEs.