Alexander Yu. Dorogov

St. Petersburg state electrotechnical university

The article proposes a method for constructing generative models based on pyramid neural networks of fast learning (FNN). The model construction is based on the probabilistic principal component method (PPCA). The PPCA method allows us to analytically construct matrices of optimal decoders that are capable to reconstructing images from random latent variables of small dimension distributed according to a normal probability law. The implementation of PPCA- decoders in the FNN class makes it possible to represent decoders in the form of series-parallel structures that provide high performance due to the structural parallelization of operations. The paper presents methods for teaching FNN to decoder matrices. Training is performed in a finite number of steps and does not require iterative procedures. Examples of constructing implementing FNN for the MNIST dataset are given. The results of generating images similar to the MNIST set are shown. A comparison with the classical variational autoencoder has been performed. The field of expedient use of generative models of PPCA is defined.

generative model, variational autoencoder, probabilistic principal component method, fast neural network, pyramid structure