Pedagogical agents as learning companions: the impact of agent emotion and gender

Abstract  The potential of emotional interaction between human and computer has recently interested researchers in human–computer interaction. The instructional impact of this interaction in learning environments has not been established, however. This study examined the impact of emotion and gender of a pedagogical agent as a learning companion (PAL) on social judgements, interest, self-efficacy, and learning. Two experiments investigated separately the effects of a PAL's emotional expression and empathetic response. Experiment 1 focused on emotional expression (positive vs. negative vs. neutral) and gender (male vs. female) with a sample of 142 male and female college students in a computer literacy course. Experiment 2 investigated the impact of empathetic response (responsive vs. non-responsive) and gender with 56 pre-service teachers. Overall, the results yielded main and interaction effects of PAL emotion and gender on the dependent variables. In particular, the PAL's empathetic response had a positive impact on learner interest and self-efficacy; PAL gender had a positive impact on recall. The findings imply that the emotion and the gender of the digital learning companion could be utilized to optimize college students' motivation and learning.     

基于PageRank的主动学习算法

在许多分类任务中,存在大量未标记的样本,并且获取样本标签耗时且昂贵。利用主动学习算法确定最应被标记的关键样本,来构建高精度分类器,可以最大限度地减少标记成本。本文提出一种基于PageRank的主动学习算法（PAL）,充分利用数据分布信息进行有效的样本选择。利用PageRank根据样本间的相似度关系依次计算邻域、分值矩阵和排名向量;选择代表样本,并根据其相似度关系构建二叉树,利用该二叉树对代表样本进行聚类,标记和预测;将代表样本作为训练集,对其他样本进行分类。实验采用8个公开数据集,与5种传统的分类算法和3种流行的主动学习算法比较,结果表明PAL算法能取得更好的分类效果。     

小样本困境下的深度学习图像识别综述

图像识别是图像研究领域的核心问题,解决图像识别问题对人脸识别、自动驾驶、机器人等各领域研究都有重要意义.目前广泛使用的基于深度神经网络的机器学习方法,已经在鸟类分类、人脸识别、日常物品分类等图像识别数据集上达到了超过人类的水平,同时越来越多的工业界应用开始考虑基于深度神经网络的方法,以完成一系列图像识别业务.但是深度学...     

一种提高同步时序PAL结构解析效率的方法

ＰＡＬ器件，尤其是高性能的ＰＡＬ２２Ｖ１０或者２２ＶＰ１０在最新的逻辑电路设计中广泛被采用。国内一些单位曾经做过ＰＡＬ解析方面的工作，但有状态探测时间长，采集数据不准确等缺陷，其解析效果不佳。本文对于同步时序结构的ＰＡＬ器件，提出了一种行之有效的解析方法，它不仅解析速度快，而且可以提高解析的准确性，给研究和分析含有加密的ＰＡＬ器件的逻辑设计者，解决了由于译码不出ＰＡＬ内部码点所带来的一系列困难，并为想要形成ＰＡＬ解析方面产品的用户提供了有效的方法。     

Evaluation of AVHRR PAL and GIMMS 10‐day composite NDVI time series products using SPOT‐4 vegetation data for the African continent

Global 8 km resolution AVHRR (advanced very high resolution radiometer) NDVI (normalized difference vegetation index) 10‐day composite data sets have been used for numerous local to global scale vegetation time series studies during recent years. AVHRR Pathfinder (PAL) NDVI was available from 1981 until 2001, and the new AVHRR GIMMS NDVI was available from 1981 to the present time. A number of aspects potentially introduce noise in the NDVI data set due to the AVHRR sensor design and data processing. NDVI from SPOT‐4 VGT data is considered an improvement over AVHRR, and for this reason it is important to examine how and if the differences in sensor design and processing influence continental scale NDVI composite products. In this study, the quality of these AVHRR NDVI time series are evaluated by the continental scale 1 km resolution SPOT‐4 vegetation (VGT) 10‐day composite (S10) NDVI data. Three years of AVHRR PAL (1998–2000) and seven years of GIMMS (1998–2004) have been compared to 8 km resampled SPOT‐4 VGT (1998–2004) data. The dynamic range of SPOT‐4 VGT NDVI tends to be higher than the AVHRR PAL NDVI, whereas there is an exact match between AVHRR GIMMS NDVI and SPOT‐4 VGT NDVI. Ortho‐regression analysis on annually integrated values of AVHRR PAL/GIMMS and SPOT‐4 VGT on a continental scale reveals high correlations amongst the AVHRR and the SPOT data set, with lowest RMSE (root mean square error) on the GIMMS/SPOT‐4 VGT compared to the PAL/SPOT‐4 VGT.

Analyses on decade data likewise show that a linear relation exists between Spot‐4 VGT NDVI and the two AVHRR composite products; GIMMS explaining most of the Spot‐4 VGT NDVI variance compared to PAL. These results show that the AVHRR GIMMS NDVI is more consistent with Spot‐4 VGT NDVI compared to AVHRR PAL versus Spot‐4 VGT NDVI (in terms of RMSE and dynamic range) and can therefore be considered the more accurate long time AVHRR data record. Analyses performed on monthly maximum composites and decade composite data, however, reveal intra‐annual variations in the correlation between SPOT‐4 VGT and the two AVHRR data sets, which are attributed to different cloud masking algorithms. The SPOT‐4 VGT cloud‐screening algorithm is insufficient, thereby suppressing the rainy season NDVI.     

KNOWLEDGE REPRESENTATION FOR LEADERSHIP STORIES

Leadership requires making decisions and implementing the results by influencing those being lead. The Personal Access to Leadership project, PAL, is constructing tools to assist leaders in creative ways and assisting the development of leaders. The tools are knowledge-based systems employing shallow understanding of the domain. The approaches used provide guidance, but do not generate solutions. One aspect that continues under exploration in PAL is the use of stories for training and guiding leaders. In order to make such support systems work, a representation is needed to enable locating useful stories related to the task of a leader. This article defines a model for story representation for the PAL tools, called the PAL tool IdeaLab. The IdeaLab is the tool to which stories are being added, as a help system to support and extend the user's thinking.     

弱监督场景下的行人重识别研究综述

近年来随着智能监控领域的不断发展,行人重识别问题逐渐受到学术界和工业界的广泛关注,其主要研究将不同摄像头下相同身份的行人图像进行关联.当前大部分研究工作关注在有监督场景下,即给定的训练数据都存在标记信息,然而考虑到数据标注工作的高成本,这在现实应用中往往是难以拓展的.本文关注于弱监督场景下的行人重识别算法,包括无监督场景和半监督场景,并且对当前先进的方法进行了分类和描述.对于无监督场景的行人重识别算法,本文根据其技术类型划分为五类,分别为基于伪标记的方法、基于图像生成的方法、基于实例分类的方法、基于领域自适应的方法和其他方法;对于半监督场景的行人重识别方法,本文根据其场景类型划分为四类,分别为少量的人有标记的场景、每一个人有少量标记的场景、基于tracklet的学习的场景和摄像头內有标记但摄像头间无标记的场景.最后,本文对当前行人重识别的相关数据集进行了整理,并对现有的弱监督方法的实验结果进行分析与总结.     

整合全局—局部度量学习的人体目标再识别

目的 人体目标再识别的任务是匹配不同摄像机在不同时间、地点拍摄的人体目标。受光照条件、背景、遮挡、视角和姿态等因素影响,不同摄相机下的同一目标表观差异较大。目前研究主要集中在特征表示和度量学习两方面。很多度量学习方法在人体目标再识别问题上了取得了较好的效果,但对于多样化的数据集,单一的全局度量很难适应差异化的特征。对此,有研究者提出了局部度量学习,但这些方法通常需要求解复杂的凸优化问题,计算繁琐。方法 利用局部度量学习思想,结合近几年提出的XQDA（cross-view quadratic discriminant analysis）和MLAPG（metric learning by accelerated proximal gradient）等全局度量学习方法,提出了一种整合全局和局部度量学习框架。利用高斯混合模型对训练样本进行聚类,在每个聚类内分别进行局部度量学习;同时在全部训练样本集上进行全局度量学习。对于测试样本,根据样本在高斯混合模型各个成分下的后验概率将局部和全局度量矩阵加权结合,作为衡量相似性的依据。特别地,对于MLAPG算法,利用样本在各个高斯成分下的后验概率,改进目标损失函数中不同样本的损失权重,进一步提高该方法的性能。结果 在VIPeR、PRID 450S和QMUL GRID数据集上的实验结果验证了提出的整合全局—局部度量学习方法的有效性。相比于XQDA和MLAPG等全局方法,在VIPeR数据集上的匹配准确率提高2.0%左右,在其他数据集上的性能也有不同程度的提高。另外,利用不同的特征表示对提出的方法进行实验验证,相比于全局方法,匹配准确率提高1.3%~3.4%左右。结论 有效地整合了全局和局部度量学习方法,既能对多种全局度量学习算法的性能做出改进,又能避免局部度量学习算法复杂的计算过程。实验结果表明,对于使用不同的特征表示,提出的整合全局—局部度量学习框架均可对全局度量学习方法做出改进。     

视频显著性检测研究进展

视频显著性检测是计算机视觉领域的一个热点研究方向,其目的在于通过联合空间和时间信息实现视频序列中与运动相关的显著性目标的连续提取.由于视频序列中目标运动模式多样、场景复杂以及存在相机运动等,使得视频显著性检测极具挑战性.本文将对现有的视频显著性检测方法进行梳理,介绍相关实验数据集,并通过实验比较分析现有方法的性能.首先,本文介绍了基于底层线索的视频显著性检测方法,主要包括基于变换分析的方法、基于稀疏表示的方法、基于信息论的方法、基于视觉先验的方法和其他方法五类.然后,对基于学习的视频显著性检测方法进行了总结,主要包括传统学习方法和深度学习方法,并着重对后一类方法进行了介绍.随后,介绍了常用的视频显著性检测数据集,给出了四种算法性能评价指标,并在不同数据集上对最新的几种算法进行了定性和定量的比较分析.最后,对视频显著性检测的关键问题进行了总结,并对未来的发展趋势进行了展望.     

模仿排序学习模型

文档排序一直是信息检索(IR)领域的关键任务之一。受益于马尔科夫决策过程强大的建模能力,以及强化学习方法强大的求解能力,近年来基于强化学习的排序模型被提出并取得了良好效果。然而,由于候选文档中会包含大量的不相关文档,导致基于“试错”的强化学习方法存在效率低下的问题。为解决上述问题,该文提出了一种基于模仿学习的排序学习算法IR-DAGGER,其基于文档标注信息构建专家策略,在保证文档排序精度的同时提高了算法的学习效率。为了测试IR-DAGGER的性能,该文基于面向相关性排序任务的OHSUMED数据集和面向多样化排序的TREC数据集进行了实验,实验结果表明IR-DAGGER在上述两个数据集上均提升了文档排序的精度和效率。     

The effects of a theory‐based summary writing tool on students' summary writing

This paper focuses on the design and evaluation of a theory‐based computer‐assisted summary writing learning environment called Summary Writing‐PAL (SW‐PAL). The SW‐PAL was developed based on four aspects: summarizing strategies, learning theories, prior knowledge, and cognitive load. A quasi‐experiment that involved 58 undergraduates majoring in Computer Science was conducted to examine the effectiveness of SW‐PAL in writing summaries. Two intact classes were selected with 28 and 30 students in control and experimental groups, respectively. The conventional teaching approach was employed in the control group, whereas the SW‐PAL was introduced to the experimental group. Pretest and posttest were administrated to both groups. The findings indicated that SW‐PAL improved students' summary writing performance. A significant variance was noted between intrinsic and extraneous cognitive load among students with varying levels of English proficiency in the experimental group, signifying that the SW‐PAL is more suitable for students with lower language proficiency.     

Dissecting learning tactics in MOOC using ordered network analysis

Background

Select and enact appropriate learning tactics that advance learning has been considered a critical set of skills to successfully complete highly flexible online courses, such as Massive open online courses (MOOCs). However, limited by analytic methods that have been used in the past, such as frequency distribution, sequence mining and process mining, we lack a deep, complete and detailed understanding of the learning tactics used by MOOC learners.

Objectives

In the present study, we proposed four major dimensions to better interpret and understand learning tactics, which are frequency, continuity, sequentiality and role of learning actions within tactics. The aim of this study was to examine to what extent can a new analytic technique, the ordered network analysis (ONA), deepen the understanding of MOOC learning tactics compared to using other methods.

Methods

In particular, we performed a fine-grained analysis of learning tactics detected from more than 4 million learning events in the behavioural trace data of 8788 learners who participated in a large-scale MOOC ‘Flipped Classroom’.

Results and Conclusions

We detected eight learning tactics, and then chose one typical tactic as an example to demonstrate how the ONA technique revealed all four dimensions and provided deeper insights into this MOOC learning tactic. Most importantly, based on the comparison with different methods such as process mining, we found that the ONA method provided a unique opportunity and novel insight into the roles of different learning actions in tactics which was neglected in the past.

Takeaway

In summary, we conclude that ONA is a promising technique that can benefit the research on learning tactics, and ultimately benefit MOOC learners by strengthening the strategic support.     

基于多核集成的在线半监督学习方法

在很多实时预测任务中,学习器需对实时采集到的数据在线地进行学习.由于数据采集的实时性,往往难以为采集到的所有数据提供标记.然而,目前的在线学习方法并不能利用未标记数据进行学习,致使学得的模型并不能即时反映数据的动态变化,降低其实时响应能力.提出一种基于多核集成的在线半监督学习方法,使得在线学习器即使在接收到没有标记的数据时也能进行在线学习.该方法采用多个定义在不同RKHS中的函数对未标记数据预测的一致程度作为正则化项,在此基础上导出了多核集成在线半监督学习的即时风险函数,然后借助在线凸规划技术进行求解.在UCl数据集上的实验结果以及在网络入侵检测上的应用表明,该方法能够有效利用数据流中未标记数据来提升在线学习的性能.     

基于深度神经网络的少样本学习综述

如何从少数训练样本中学习并识别新的类别对于深度神经网络来说是一个具有挑战性的问题。针对如何解决少样本学习的问题,全面总结了现有的基于深度神经网络的少样本学习方法,涵盖了方法所用模型、数据集及评估结果等各个方面。具体地,针对基于深度神经网络的少样本学习方法,提出将其分为数据增强方法、迁移学习方法、度量学习方法和元学习方法四种类别;对于每个类别,进一步将其分为几个子类别,并且在每个类别与方法之间进行一系列比较,以显示各种方法的优劣和各自的特点。最后强调了现有方法的局限性,并指出了少样本学习研究领域未来的研究方向。     

Acrobot control by learning the switching of multiple controllers

Reinforcement learning (RL) has been applied to constructing controllers for nonlinear systems in recent years. Since RL methods do not require an exact dynamics model of the controlled object, they have a higher flexibility and potential for adaptation to uncertain or nonstationary environments than methods based on traditional control theory. If the target system has a continuous state space whose dynamic characteristics are nonlinear, however, RL methods often suffer from unstable learning processes. For this reason, it is difficult to apply RL methods to control tasks in the real world. In order to overcome the disadvantage of RL methods, we propose an RL scheme combining multiple controllers, each of which is constructed based on traditional control theory. We then apply it to a swinging-up and stabilizing task of an acrobot with a limited torque, which is a typical but difficult task in the field of nonlinear control theory. Our simulation result showed that our method was able to realize stable learning and to achieve fairly good control.This work was presented, in part, at the 9th International Symposium on Artificial Life and Robotics, Oita, Japan, January 28–30, 2004     

数据异常的监测技术综述

在目前大数据的环境下,相对于正常数据,异常类数据更难获取,也显得更加重要。异常检测的目的是检测出异于正常主体的活动数据。异常检测适用于机器故障诊断、数据挖掘以及疾病和入侵检测等多个领域。基于目前大量的异常检测方法,主要从异常类数据的有无来阐述,根据这个框架将主要的异常检测方法进行了分类,并评价了这些方法的优劣;最后重点讨论了基于深度学习的大数据异常检测方法,并分别介绍了不同的方法及相关的应用和未来的研究热点。