2017/05/19 by Luca Ambrogioni, Ambrogioni, Luca, Umut Güçlü +5 · 4 citations
Computer Science · Mathematics · #Bayesian Methods and Mixture Models #FOS: Computer and information sciences #Gaussian Processes and Bayesian Inference #Generative Adversarial Networks and Image Synthesis #Human Pose and Action Recognition #Machine Learning (stat.ML) #Morphological variations and asymmetry
paper · pdf · doi:10.48550/arxiv.1705.07111
openalex publication_date 2017/05/19 · openalex created_date 2022/09/30 · openalex updated_date 2026/07/28
This paper introduces the kernel mixture network, a new method for\nnonparametric estimation of conditional probability densities using neural\nnetworks. We model arbitrarily complex conditional densities as linear\ncombinations of a family of kernel functions centered at a subset of training\npoints. The weights are determined by the outer layer of a deep neural network,\ntrained by minimizing the negative log likelihood. This generalizes the popular\nquantized softmax approach, which can be seen as a kernel mixture network with\nsquare and non-overlapping kernels. We test the performance of our method on\ntwo important applications, namely Bayesian filtering and generative modeling.\nIn the Bayesian filtering example, we show that the method can be used to\nfilter complex nonlinear and non-Gaussian signals defined on manifolds. The\nresulting kernel mixture network filter outperforms both the quantized softmax\nfilter and the extended Kalman filter in terms of model likelihood. Finally,\nour experiments on generative models show that, given the same architecture,\nthe kernel mixture network leads to higher test set likelihood, less\noverfitting and more diversified and realistic generated samples than the\nquantized softmax approach.\n