Statistical Methodology towards Technology

点击翻译按钮获取中文摘要

This proposal puts in place two main long-term research directions to use statistical methodologies towards artificial intelligence. These two directions have the potential of reviving statistical models as the main tool for complicated tasks such as object detection, image classification, and machine translation. Direction 1: convolutional probabilistic models. This direction allows to provide a deep and richer version of statistical mixture models combined with product partition models. This direction allows to develop a statistically meaningful alternative to deep neural networks with more advantages. These advantages are significant thank to the well-established theory of mixture models and product partition models. Direction 2: edge computation of deep neural networks. This direction allows to bring smart technology to edge devices such as cell phone and wearables. Currently most of smart computations rely on cloud services and eats data bandwidth. This research proposal allows to explore avenues for using well-established penalization and regularization methodologies, to bring computations on edge devices.