Improving Generative and Discriminative Modelling Performance by Implementing Learning Constraints in Encapsulated Variational Autoencoders

摘要

Learning latent representations of observed data that can favour both discriminative and generative tasks remains a challenging task in artificial-intelligence (AI) research. Previous attempts that ranged from the convex binding of discriminative and generative models to the semisupervised learning paradigm could hardly yield optimal performance on both generative and discriminative tasks. To this end, in this research, we harness the power of two neuroscience-inspired learning constraints, that is, dependence minimisation and regularisation constraints, to improve generative and discriminative modelling performance of a deep generative model. To demonstrate the usage of these learning constraints, we introduce a novel deep generative model: encapsulated variational autoencoders (EVAEs) to stack two different variational autoencoders together with their learning algorithm. Using the MNIST digits dataset as a demonstration, the generative modelling performance of EVAEs was improved with the imposed dependence-minimisation constraint, encouraging our derived deep generative model to produce various patterns of MNIST-like digits. Using CIFAR-10(4K) as an example, a semisupervised EVAE with an imposed regularisation learning constraint was able to achieve competitive discriminative performance on the classification benchmark, even in the face of state-of-the-art semisupervised learning approaches.