Dec, 2016

深度学习的概率框架

TL;DRDeveloped a probabilistic framework for deep learning based on the Deep Rendering Mixture Model (DRMM), which captures variations in data due to latent task nuisance variables and provides a principled route to the improvement of deep convolutional neural networks (DCNs). The DRMM is a powerful alternative to DCN back-propagation, leading to faster training and achieving superior accuracy in supervised digit classification along with state-of-the-art results on the MNIST benchmark and competitive results on the CIFAR10 benchmark in semi-supervised and unsupervised learning tasks.