Context-aware Mixture of Deep Neural Networks

Abstract

Alpha-beta network is a mixture of deep neural networks, implementing a mixture of experts, where each component is a neural network. It is trained using the expectation-maximization algorithm. It enables context-awareness as each component is pushed to give context-specific predictions. Such structure enables context uncertainty quantification as well. The effectiveness of alpha-beta network was assessed using two real-world activity datasets: UCI OPPORTUNITY and an in-house dataset. The model has shown superior performance compared to the baselines.