Stochastic Variational Autoencoder

This work explores a novel generative modeling approach inspired by variational autoencoders (VAEs). Traditional VAEs rely on a recognition model (encoder) that approximates the latent posterior with a single gaussian distribution for each input, limiting their flexibility in capturing complex data distributions. In contrast, we propose a modified recognition model that utilizes stochastic mixtures of gaussians, allowing for a more expressive latent representation. By leveraging stochastic neural networks within the VAE framework, we aim to achieve a tighter evidence lower bound (ELBO) on the log-likelihood of the data. Our approach is a preliminary investigation to enhance the latent space structure and improve generative performance by incorporating richer uncertainty modeling.