Deconvolutional Density Network: Modeling Free-Form Conditional Distributions

Conditional density estimation (CDE) is the task of estimating probability an event conditioned on some inputs. A neural network (NN) can also be used to compute output distribution for continuous-domain, which viewed as extension regression task. Nevertheless, it difficult explicitly approximate a without knowing information its general form priori. In order fit arbitrary conditional distribution, discretizing continuous domain into bins effective strategy, long we have sufficiently narrow and very large data. However, collecting enough data often hard reach falls far short that ideal in many circumstances, especially multivariate CDE curse dimensionality. this paper, demonstrate benefits modeling free-form distributions using deconvolution-based net framework, coping with deficiency problems discretization. It has advantage being flexible but takes hierarchical smoothness offered by deconvolution layers. We compare our method number other density-estimation approaches show Deconvolutional Density Network (DDN) outperforms competing methods univariate tasks. The code DDN available at https://github.com/NBICLAB/DDN.