目标不同视角下观察信息的迁移和再利用

基于视觉信息的目标检测和识别模型在训练时往往依赖于来自于训练样本的视角信息,然而附带了视角信息的训练样本通常只有很少的数据库可以提供。当此类信息缺失时,传统的通用目标检测系统通常通过一些非监督学习方法来对样本的视角信息进行粗略估计。本文改进并引入了一种选择性迁移学习方法即TransferBoost方法来解决目标视角信息缺失的问题。本文TransferBoost方法基于GentleBoost框架实现,该方法通过重新利用其它类别样本中的先验信息来提升当前类别样本的学习质量。当给定一个标定完善的样本集作为源数据库时,TransferBoost通过同时调整每个样本的权值和每个源任务的权值实现样本级和任务级的两级知识迁移。这种双层迁移学习更有效地从混合了相关源数据和不相关源数据的数据集中提取了有用的信息。实验结果表明,和直接使用传统的机器学习方法相比较,迁移学习方法所需要的训练样本数大大减少,从而降低了目标检测与识别系统的训练代价,扩展了现有系统的应用范围。     

基于迁移学习的讽刺检测

准确的讽刺检测对于情感分析等任务至关重要.传统的方法严重依赖于离散的人工制定的特征.现有的研究大多将讽刺检测作为一种标准的监督学习文本分类任务,但是监督学习需要有大量数据,而这些数据的收集和标注都存在困难.由于目标任务有限的数据集可能导致讽刺检测的低性能,为此将讽刺检测作为一种迁移学习任务,将讽刺标记文本的监督学习与外部分析资源的知识转移相结合.通过转移的资源知识来改进神经网络模型,以此提升对目标任务的检测性能.在公开可用的数据集上的实验结果表明,提出的基于迁移学习的讽刺检测模型优于现有较先进的讽刺检测模型.     

Estimating variable structure and dependence in multitask learning via gradients

We consider the problem of hierarchical or multitask modeling where we simultaneously learn the regression function and the underlying geometry and dependence between variables. We demonstrate how the gradients of the multiple related regression functions over the tasks allow for dimension reduction and inference of dependencies across tasks jointly and for each task individually. We provide Tikhonov regularization algorithms for both classification and regression that are efficient and robust for high-dimensional data, and a mechanism for incorporating a priori knowledge of task (dis)similarity into this framework. The utility of this method is illustrated on simulated and real data.     

结合记忆与迁移学习的小样本学习

小样本学习是视觉识别中的一个受关注的领域,旨在通过少量的数据来学习新的视觉概念。为了解决小样本问题,一些元学习方法提出从大量辅助任务中学习可迁移的知识并将其应用于目标任务上。为了更好地对知识进行迁移,提出了一种基于记忆的迁移学习方法。提出一种权重分解策略,将部分权重分解为冻结权重与可学习权重,在迁移学习中通过固定冻结权重,仅更新可学习权重的方式来减少模型需要学习的参数。通过一个额外的记忆模块来存储之前任务的经验,在学习新任务时,这些经验被用来初始化模型的参数状态,以此更好地进行迁移学习。通过在miniImageNet、tieredImageNet以及CUB数据集上的实验结果表明,相对于其他先进的方法,该方法在小样本分类任务上取得了具有竞争力甚至是更好的表现。     

Multitask Learning

Multitask Learning is an approach to inductive transfer that improves generalization by using the domain information contained in the training signals of related tasks as an inductive bias. It does this by learning tasks in parallel while using a shared representation; what is learned for each task can help other tasks be learned better. This paper reviews prior work on MTL, presents new evidence that MTL in backprop nets discovers task relatedness without the need of supervisory signals, and presents new results for MTL with k-nearest neighbor and kernel regression. In this paper we demonstrate multitask learning in three domains. We explain how multitask learning works, and show that there are many opportunities for multitask learning in real domains. We present an algorithm and results for multitask learning with case-based methods like k-nearest neighbor and kernel regression, and sketch an algorithm for multitask learning in decision trees. Because multitask learning works, can be applied to many different kinds of domains, and can be used with different learning algorithms, we conjecture there will be many opportunities for its use on real-world problems.     

基于共享BERT和门控多任务学习的事件检测方法

事件检测任务的目标是从文本中自动获取结构化的事件信息。目前基于表示学习的神经事件检测方法能够有效利用潜在语义信息,但人工标注数据集的语义知识含量有限,制约了神经网络模型的认知广度。相对地,多任务表示学习框架,有助于模型同时学习不同任务场景中的语义知识,从而提升其认知广度。BERT预训练模型得益于大规模语言资源的充沛语义信息,具有高适应性(适应不同任务)的语义编码能力。因此,该文提出了一种基于BERT的多任务事件检测模型。该方法将BERT已经包含的语义知识作为基础,进一步提升多任务模型的表示、学习和语义感知能力。实验表明,该方法有效提高了事件检测的综合性能,其在ACE2005语料集上事件分类的F₁值达到了76.7%。此外,该文在实验部分对多任务模型的训练过程进行了详解,从可解释性的层面分析了多任务架构对事件检测过程的影响。     

一种面向多源领域的实例迁移学习

在迁移学习最大的特点就是利用相关领域的知识来帮助完成目标领域中的学习任务,它能够有效地在相似的领域或任务之间进行信息的共享和迁移,使传统的从零开始的学习变成可积累的学习,具有成本低、效率高等优点.针对源领域数据和目标领域数据分布类似的情况,提出一种基于多源动态TrAdaBoost的实例迁移学习方法.该方法考虑多个源领域知识,使得目标任务的学习可以充分利用所有源领域信息,每次训练候选分类器时,所有源领域样本都参与学习,可以获得有利于目标任务学习的有用信息,从而避免负迁移的产生.理论分析验证了所提算法较单源迁移的优势,以及加入动态因子改善了源权重收敛导致的权重熵由源样本转移到目标样本的问题.实验结果验证了此算法在提高识别率方面的优势.     

交通图像中机动车联合检测与识别

目的 机动车检测和属性识别是智能交通系统中的基本任务,现有的方法通常将检测和识别分开进行,导致以下两个问题：一是检测算法与识别任务在时序上存在耦合问题,增加了算法设计的复杂度;二是多个任务模块及其交互会增加计算负载,降低了智能交通系统的执行效率。为了解决以上问题,结合机动车辆视觉属性与检测之间的联系,提出机动车联合检测与识别方法,将检测和属性识别任务整合在一个算法框架中完成。方法 首先,将车辆颜色与类型融合到检测算法中,使用多任务学习框架对机动车的属性识别任务与定位任务建模,在检测的同时完成属性识别。进一步地,针对智能交通系统中数据分布不均匀、呈现长尾现象的问题,将多任务学习框架与在线难例挖掘算法相结合,降低该现象给模型优化带来的危害。结果 为了验证本文提出的方法,构建了拥有12 712幅图像,包含19 398辆机动车的道路车辆图像数据集。在该数据集上,使用机动车联合检测与识别算法取得了85.6%的检测精度,优于SSD （single shot detector）与Faster-RCNN检测方法。针对识别任务,本文方法对于颜色与类型属性的识别准确率分别达到了91.3%和91.8%。结论 车辆颜色和类型作为机动车的重要视觉特征,综合利用以上线索有助于提高机动车检测的效果,同时能够得到良好的属性识别性能。除此之外,使用一个高度集成的框架完成多个任务,可以提升智能交通系统的运行效率。     

联合多任务学习的人脸超分辨率重建

目的 人脸超分辨率重建是特定应用领域的超分辨率问题,为了充分利用面部先验知识,提出一种基于多任务联合学习的深度人脸超分辨率重建算法。方法 首先使用残差学习和对称式跨层连接网络提取低分辨率人脸的多层次特征,根据不同任务的学习难易程度设置损失权重和损失阈值,对网络进行多属性联合学习训练。然后使用感知损失函数衡量HR（high-resolution）图像与SR（super-resolution）图像在语义层面的差距,并论证感知损失在提高人脸语义信息重建效果方面的有效性。最后对人脸属性数据集进行增强,在此基础上进行联合多任务学习,以获得视觉感知效果更加真实的超分辨率结果。结果 使用峰值信噪比（PSNR）和结构相似度（SSIM）两个客观评价标准对实验结果进行评价,并与其他主流方法进行对比。实验结果显示,在人脸属性数据集（CelebA）上,在放大8倍时,与通用超分辨率MemNet（persistent memory network）算法和人脸超分辨率FSRNet（end-to-end learning face super-resolution network）算法相比,本文算法的PSNR分别提升约2.15 dB和1.2 dB。结论 实验数据与效果图表明本文算法可以更好地利用人脸先验知识,产生在视觉感知上更加真实和清晰的人脸边缘和纹理细节。     

基于联合学习的生物医学因果关系抽取

生物医学因果关系抽取是BioCreative社区提出的一项评测任务,旨在挖掘生物医学实体间丰富的语义关系,并用生物医学表征语言(biological expression language, BEL)来表示。与传统的实体关系抽取不同,该任务不仅包含实体间因果关系的抽取,还包含实体功能的识别。此前已经提出了一些该任务的解决方法,但均未考虑这两个子任务间的关联性。该文基于多任务的思想,提出一种二元关系抽取和一元功能识别共同决策的联合学习模式。首先两个任务共享底层向量表示,然后利用长短期记忆(long short-term memory, LSTM)网络和门控机制学习两个任务之间的交互表示,最后分别进行分类预测。实验结果表明,该方法能够融合两个子任务的信息,在2015 BC-V测试集上获得了45.3%的F值。     

Learning with few examples for binary and multiclass classification using regularization of randomized trees

The human visual system is often able to learn to recognize difficult object categories from only a single view, whereas automatic object recognition with few training examples is still a challenging task. This is mainly due to the human ability to transfer knowledge from related classes. Therefore, an extension to Randomized Decision Trees is introduced for learning with very few examples by exploiting interclass relationships. The approach consists of a maximum a posteriori estimation of classifier parameters using a prior distribution learned from similar object categories. Experiments on binary and multiclass classification tasks show significant performance gains     

A notion of task relatedness yielding provable multiple-task learning guarantees

The approach of learning multiple “related” tasks simultaneously has proven quite successful in practice; however, theoretical justification for this success has remained elusive. The starting point for previous work on multiple task learning has been that the tasks to be learned jointly are somehow “algorithmically related”, in the sense that the results of applying a specific learning algorithm to these tasks are assumed to be similar. We offer an alternative approach, defining relatedness of tasks on the basis of similarity between the example generating distributions that underlie these tasks. We provide a formal framework for this notion of task relatedness, which captures a sub-domain of the wide scope of issues in which one may apply a multiple task learning approach. Our notion of task similarity is relevant to a variety of real life multitask learning scenarios and allows the formal derivation of generalization bounds that are strictly stronger than the previously known bounds for both the learning-to-learn and the multitask learning scenarios. We give precise conditions under which our bounds guarantee generalization on the basis of smaller sample sizes than the standard single-task approach. Editors: Daniel Silver, Kristin Bennett, Richard Caruana. A preliminary version of this paper appears in the proceedings of COLT’03, (Ben-David and Schuller 2003).     

注意力迁移的联合平衡领域自适应

目的 现有的图像识别方法应用于从同一分布中提取的训练数据和测试数据时具有良好性能,但这些方法在实际场景中并不适用,从而导致识别精度降低。使用领域自适应方法是解决此类问题的有效途径,领域自适应方法旨在解决来自两个领域相关但分布不同的数据问题。方法 通过对数据分布的分析,提出一种基于注意力迁移的联合平衡自适应方法,将源域有标签数据中提取的图像特征迁移至无标签的目标域。首先,使用注意力迁移机制将有标签源域数据的空间类别信息迁移至无标签的目标域。通过定义卷积神经网络的注意力,使用关注信息来提高图像识别精度。其次,基于目标数据集引入网络参数的先验分布,并且赋予网络自动调整每个领域对齐层特征对齐的能力。最后,通过跨域偏差来描述特定领域的特征对齐层的输入分布,定量地表示每层学习到的领域适应性程度。结果 该方法在数据集Office-31上平均识别准确率为77.6%,在数据集Office-Caltech上平均识别准确率为90.7%,不仅大幅领先于传统手工特征方法,而且取得了与目前最优的方法相当的识别性能。结论 注意力迁移的联合平衡领域自适应方法不仅可以获得较高的识别精度,而且能够自动学习领域间特征的对齐程度,同时也验证了进行域间特征迁移可以提高网络优化效果这一结论。     

Task-dependent processing of tables and graphs

In two experiments participants had to detect changes in periodic sinusoidal functions, displayed in either graphic or tabular displays. Graphs had a major advantage over tables when the task required considering configurations of data. Both displays led to similar results when task performance could rely on inspecting individual data points. With graphs almost all participants reported using the optimal method for detecting changes in the function, i.e., they used the method requiring the least effort to perform the task. With tables only about half used the optimal detection method, and these participants showed transfer of learning of detection methods between tasks. Experience in using a detection method led to improved performance if the new task relied on the same method of detection. These findings demonstrate the need to consider task performance methods when determining the relative value of different displays. The set of tasks for which a display is used is likely to affect performance and needs to be analysed as a whole, since methods employed for one task can affect task performance in other tasks.     

Efficient Exploration for Multi-Agent Reinforcement Learning via Transferable Successor Features

In multi-agent reinforcement learning (MARL), the behaviors of each agent can influence the learning of others, and the agents have to search in an exponentially enlarged joint-action space. Hence, it is challenging for the multi-agent teams to explore in the environment. Agents may achieve suboptimal policies and fail to solve some complex tasks. To improve the exploring efficiency as well as the performance of MARL tasks, in this paper, we propose a new approach by transferring the knowledge across tasks. Differently from the traditional MARL algorithms, we first assume that the reward functions can be computed by linear combinations of a shared feature function and a set of task-specific weights. Then, we define a set of basic MARL tasks in the source domain and pre-train them as the basic knowledge for further use. Finally, once the weights for target tasks are available, it will be easier to get a well-performed policy to explore in the target domain. Hence, the learning process of agents for target tasks is speeded up by taking full use of the basic knowledge that was learned previously. We evaluate the proposed algorithm on two challenging MARL tasks: cooperative box-pushing and non-monotonic predator-prey. The experiment results have demonstrated the improved performance compared with state-of-the-art MARL algorithms.      

基于迁移学习的多任务分配算法

对于多任务分配问题，传统的方法针对每一个任务独立地寻找一个最优分配方案，没有考虑任务间的关联以及历史经验对新任务分配的影响，因而复杂度较高。研究了多智能体系统中的多任务分配问题，通过迁移学习来加速任务分配以及子任务的完成。在分配目标任务时，通过计算当前任务和历史任务的相似度找到最适合的源任务，再将源任务的分配模式迁移到目标任务中，并在完成子任务的过程中使用迁移学习，从而提高效率，节约时间。最后，通过“格子世界”的实验证明了该算法在运行时间和平均带折扣回报方面都优于基于Q学习的任务分配算法。     

Real-time video fire smoke detection by utilizing spatial-temporal ConvNet features

Fire is one of the most dangerous disasters threatening human life and property globally. In order to reduce fire losses, researches on video analysis for early smoke detection have become particularly significant. However, it is still a challenging task to extract stable features for smoke recognition, largely due to its variations in color, shapes and texture. Classical convolutional neural networks can automatically learn feature representations of appearance from a single frame but fail to capture motion information between frames. For addressing this issue, in this paper, we propose a spatial-temporal based convolutional neural network for video smoke detection, and for real-time detection, propose an enhanced architecture, which utilizes a multitask learning strategy to jointly recognize smoke and estimate optical flow, capturing intra-frame appearance features and inter-frame motion features simultaneously. The effectiveness and efficiency of our proposed method is validated by experiments carried out on our self-created dataset, which achieves 97.0% detection rate and 3.5% false alarm rate with processing time of 5ms per frame, obviously outperforming existing methods.     

持续学习改进的人脸表情识别

目的 大量标注数据和深度学习方法极大地提升了图像识别性能。然而,表情识别的标注数据缺乏,训练出的深度模型极易过拟合,研究表明使用人脸识别的预训练网络可以缓解这一问题。但是预训练的人脸网络可能会保留大量身份信息,不利于表情识别。本文探究如何有效利用人脸识别的预训练网络来提升表情识别的性能。方法 本文引入持续学习的思想,利用人脸识别和表情识别之间的联系来指导表情识别。方法指出网络中对人脸识别整体损失函数的下降贡献最大的参数与捕获人脸公共特征相关,对表情识别来说为重要参数,能够帮助感知面部特征。该方法由两个阶段组成：首先训练一个人脸识别网络,同时计算并记录网络中每个参数的重要性;然后利用预训练的模型进行表情识别的训练,同时通过限制重要参数的变化来保留模型对于面部特征的强大感知能力,另外非重要参数能够以较大的幅度变化,从而学习更多表情特有的信息。这种方法称之为参数重要性正则。结果 该方法在RAF-DB（real-world affective faces database）,CK+（the extended Cohn-Kanade database）和Oulu-CASIA这3个数据集上进行了实验评估。在主流数据集RAF-DB上,该方法达到了88.04%的精度,相比于直接用预训练网络微调的方法提升了1.83%。其他数据集的实验结果也表明了该方法的有效性。结论 提出的参数重要性正则,通过利用人脸识别和表情识别之间的联系,充分发挥人脸识别预训练模型的作用,使得表情识别模型更加鲁棒。     

Transfer dimensionality reduction by Gaussian process in parallel

Dimensionality reduction has been considered as one of the most significant tools for data analysis. In general, supervised information is helpful for dimensionality reduction. However, in typical real applications, supervised information in multiple source tasks may be available, while the data of the target task are unlabeled. An interesting problem of how to guide the dimensionality reduction for the unlabeled target data by exploiting useful knowledge, such as label information, from multiple source tasks arises in such a scenario. In this paper, we propose a new method for dimensionality reduction in the transfer learning setting. Unlike traditional paradigms where the useful knowledge from multiple source tasks is transferred through distance metric, we attempt to learn a more informative mapping function between the original data and the reduced data by Gaussian process that behaves more appropriately than other parametric regression methods due to its less parametric characteristic. In our proposal, we firstly convert the dimensionality reduction problem into integral regression problems in parallel. Gaussian process is then employed to learn the underlying relationship between the original data and the reduced data. Such a relationship can be appropriately transferred to the target task by exploiting the prediction ability of the Gaussian process model and inventing different kinds of regularizers. Extensive experiments on both synthetic and real data sets show the effectiveness of our method.     

Inductive transfer for learning Bayesian networks

In several domains it is common to have data from different, but closely related problems. For instance, in manufacturing, many products follow the same industrial process but with different conditions; or in industrial diagnosis, where there is equipment with similar specifications. In these cases it is common to have plenty of data for some scenarios but very little for others. In order to learn accurate models for rare cases, it is desirable to use data and knowledge from similar cases; a technique known as transfer learning. In this paper we propose an inductive transfer learning method for Bayesian networks, that considers both structure and parameter learning. For structure learning we use conditional independence tests, by combining measures from the target task with those obtained from one or more auxiliary tasks, using a novel weighted sum of the conditional independence measures. For parameter learning, we propose two variants of the linear pool for probability aggregation, combining the probability estimates from the target task with those from the auxiliary tasks. To validate our approach, we used three Bayesian networks models that are commonly used for evaluating learning techniques, and generated variants of each model by changing the structure as well as the parameters. We then learned one of the variants with a small dataset and combined it with information from the other variants. The experimental results show a significant improvement in terms of structure and parameters when we transfer knowledge from similar tasks. We also evaluated the method with real-world data from a manufacturing process considering several products, obtaining an improvement in terms of log-likelihood between the data and the model when we do transfer learning from related products.