Improving the appeal of variational approximations
Date: 12 March 2021, Friday
Time: 4pm AEDT
Speaker: Dr Luca Maestrini (University of Technology Sydney)
Abstract:
Variational approximations facilitate approximate inference for the parameters of a variety of statistical models. However, they are sometimes criticized for being hard to implement, hindered by the sizes of model design matrices and potentially inaccurate. First, we show how the notion of variational message passing on factor graph fragments allows for repeated use of algorithmic primitives, representing enormous savings in terms of algebra and computer coding. We illustrate this concept on applications that can be modelled as inverse problems. We also explain how streamlined solutions to sparse matrix problems can be used for making fast variational inference for models with a high number of random effects and provide an illustration for multilevel models with penalized regression coefficients. Last, we show a simple remedy to reduce inaccuracy of variational approximations through an example concerning structural equation models.
Link: https://au.bbcollab.com/guest/fcf219c74ac743e89565a9e6e8d349a9
Time: 4pm AEDT
Speaker: Dr Luca Maestrini (University of Technology Sydney)
Abstract:
Variational approximations facilitate approximate inference for the parameters of a variety of statistical models. However, they are sometimes criticized for being hard to implement, hindered by the sizes of model design matrices and potentially inaccurate. First, we show how the notion of variational message passing on factor graph fragments allows for repeated use of algorithmic primitives, representing enormous savings in terms of algebra and computer coding. We illustrate this concept on applications that can be modelled as inverse problems. We also explain how streamlined solutions to sparse matrix problems can be used for making fast variational inference for models with a high number of random effects and provide an illustration for multilevel models with penalized regression coefficients. Last, we show a simple remedy to reduce inaccuracy of variational approximations through an example concerning structural equation models.
Link: https://au.bbcollab.com/guest/fcf219c74ac743e89565a9e6e8d349a9