专栏名称: 机器学习研究会

机器学习研究会是北京大学大数据与机器学习创新中心旗下的学生组织，旨在构建一个机器学习从事者交流的平台。除了及时分享领域资讯外，协会还会举办各种业界巨头/学术神牛讲座、学术大牛沙龙分享会、real data 创新竞赛等活动。

【学习】机器学习初学者教程：贝叶斯方法

机器学习研究会 · 公众号 · AI · 2017-06-14 20:52

正文

点击上方 “机器学习研究会” 可以订阅哦

摘要

转自：网路冷眼

INTRODUCTION

This tutorial is intended for readers who are interested in applying Bayesian methods to machine learning. Our intention is to teach you how to train your first Bayesian neural network, and provide a Bayesian companion to the well known getting started example in TensorFlow.

So why do we need Bayesian neural networks? Traditionally neural networks are trained to produce a point estimate of some variable of interest. For example, we might train a neural network to produce a prediction of a stock price at a future point in time using historical data. The limitation of a single point estimate is that it does not provide us with any measure of the uncertainty in this prediction. If the network predicts that the stock will increase in value with 95% confidence then we probably have an easy decision to buy, but what if the network has only a 50% confidence? With point estimates we just don't know how uncertain we are. By contrast, Bayesian neural networks enable us to estimate the uncertainty in our predictions using Bayes' rule.