A0983
Title: Probabilistic activation functions and semiparametric mean field variational learning in Bayesian neural networks
Authors: Mingwei Lin - London School of Economics and Political Science (United Kingdom) [presenting]
Giulia Livieri - The London School of Economics and Political Science (United Kingdom)
Luca Maestrini - The Australian National University (Australia)
Mauro Bernardi - University of Padova (Italy)
Abstract: Probabilistic representations for activation functions in Bayesian neural networks are proposed through the introduction of augmented variables. These models transcend traditional conjugate frameworks, which typically present intractability issues. To address these challenges, the semiparametric mean field variational approximations are implemented to manage the intractable posteriors in the parameter learning processes. Compared to Markov chain Monte Carlo (MCMC) methods, this approach maintains good approximation results and offers faster training processes. Additionally, a mixture of expert frameworks is introduced in variational learning, which further enhances prediction accuracy and holds substantial potential for reducing computational costs.
