This shows a Variational Autoencoder using Convolutional Neural Networks (CNNs) to process 3D data.
The Encoder uses convolutional layers to extract features. After the final layer, these features are 'flattened' into a single vector. This high-dimensional vector is then dramatically compressed into the few dimensions of the Latent Space.
The Decoder reverses the process, reconstructing the shape from this compact latent representation.