人工智能培训

搜索

论文代码开源:VIZDOOM和MONTEZUMAS REVENGE的专家增强演员评论家(Expert-augmented actor-critic for Vi

[复制链接]
admin 发表于 2018-9-15 10:38:20 | 显示全部楼层 |阅读模式
admin 2018-9-15 10:38:20 849 0 显示全部楼层
人工智能论文代码开源:VIZDOOM和MONTEZUMAS REVENGE的专家增强演员评论家(Expert-augmented actor-critic for ViZDoom and Montezumas Revenge)请注意该人工智能论文代码开源在github,大部分是python写的,框架可能是tensorflow或者pytorch。标签信息在监督的高光谱图像分类问题中起着重要作用。然而,目前的分类方法都忽略了一个重要且不可避免的问题---标签可能被破坏,收集训练样本的清洁标签很困难,而且往往不切实际。因此,如何从带有噪声标签的数据库中学习是一个非常实用的问题。在本文中,我们研究了标签噪声对高光谱图像分类的影响,并开发了一种随机标签传播算法(RLPA)来清除标签噪声。 RLPA的关键思想是从观察到的高光谱图像中利用知识(例如,基于超像素的光谱空间约束)并将其应用于标签传播过程。具体而言,RLPA首先构建同时考虑的谱空间概率传递矩阵(SSPTM)。基于光谱相似度和超像素的空间信息。然后,它随机选择一些训练样本作为“干净”样本,并将其余样本设置为未标记样本,并使用SSPTM将标记信息从“干净”样本传播到其余未标记样本。通过重复随机分配(“清洁”标记样本和未标记样本)和传播,我们可以为每个训练样本获得多个标签。因此,最终传播的标签可以通过多数投票算法计算。实验研究表明,RLPA可以降低噪声标签的水平,并证明了我们提出的方法优于四个主要分类器的优势,具有显着的余量 - 平均OA,AA,Kappa令人印象深刻,例如,9.18%,9.58% ,和0.1043。 Matlab源代码可通过此https URL获得
Label information plays an important role in supervised hyperspectral imageclassification problem.However, current classification methods all ignore animportant and inevitable problem---labels may be corrupted and collecting cleanlabels for training samples is difficult, and often impractical.Therefore, howto learn from the database with noisy labels is a problem of great practicalimportance.In this paper, we study the influence of label noise onhyperspectral image classification, and develop a random label propagationalgorithm (RLPA) to cleanse the label noise.The key idea of RLPA is to exploitknowledge (eg, the superpixel based spectral-spatial constraints) from theobserved hyperspectral images and apply it to the process of label propagation.Specifically, RLPA first constructs a spectral-spatial probability transfermatrix (SSPTM) that simultaneously considersthe spectral similarity andsuperpixel based spatial information.It then randomly chooses some trainingsamples as "clean" samples and sets the rest as unlabeled samples, andpropagates the label information from the "clean" samples to the rest unlabeledsamples with the SSPTM.By repeating the random assignment (of "clean" labeledsamples and unlabeled samples) and propagation, we can obtain multiple labelsfor each training sample.Therefore, the final propagated label can becalculated by a majority vote algorithm.Experimental studies show that RLPAcan reduce the level of noisy label and demonstrates the advantages of ourproposed method over four major classifiers with a significant margin---thegains in terms of the average OA, AA, Kappa are impressive, eg, 9.18%, 9.58%,and 0.1043.The Matlab source code is available atthis https URL论文代码开源:VIZDOOM和MONTEZUMAS REVENGE的专家增强演员评论家(Expert-augmented actor-critic for ViZDoom and Montezumas Revenge) e55b1gb4Qvq404iE.jpg
URL地址:https://arxiv.org/abs/1809.04212v1     ----pdf下载地址:https://arxiv.org/pdf/1809.04212v1    ----         ----github下载地址:https://github.com/junjun-jiang/RLPA    ----    论文代码开源:VIZDOOM和MONTEZUMAS REVENGE的专家增强演员评论家(Expert-augmented actor-critic for ViZDoom and Montezumas Revenge)请注意该人工智能论文代码开源在github,大部分是python写的,框架可能是tensorflow或者pytorch,keras,至于具体是哪一个没有完全测试。
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

admin当前离线
管理员

查看:849 | 回复:0

快速回复 返回顶部 返回列表