Conditional Invertible Neural Network
A variant of normalizing flows that maps a simple base distribution to a target posterior distribution through invertible transformations. It is used for parameter posterior estimation and allows exact computation of the log-likelihood of the posterior density.
Timeline (last 90 days)
The learned summary statistics are fed into two parallel branches: the model selection network computes posterior model probabilities, and multiple model-specific conditional invertible neural networks (cINNs) separately estimate the parameter posterior distributions under each model.
A conditional invertible neural network (cINN) is a variant of normalizing flows that maps a simple base distribution to the target posterior distribution through invertible transformations, and can give the log-likelihood of the posterior density in exact form.
The invertibility of cINNs constrains the structure of each network layer, requiring carefully designed coupling layers, and training is prone to ill-conditioned gradients, placing higher demands on network capacity and training stability.
All Facts (3)
The learned summary statistics are fed into two parallel branches: the model selection network computes posterior model probabilities, and multiple model-specific conditional invertible neural networks (cINNs) separately estimate the parameter posterior distributions under each model.
50%UnverifiedA conditional invertible neural network (cINN) is a variant of normalizing flows that maps a simple base distribution to the target posterior distribution through invertible transformations, and can give the log-likelihood of the posterior density in exact form.
50%UnverifiedThe invertibility of cINNs constrains the structure of each network layer, requiring carefully designed coupling layers, and training is prone to ill-conditioned gradients, placing higher demands on network capacity and training stability.
50%