基于潜层向量对齐的持续零样本学习算法

在持续学习多任务过程中,持续零样本学习旨在积累已见类知识,并用于识别未见类样本.然而,在连续学习过程中容易产生灾难性遗忘,因此,文中提出基于潜层向量对齐的持续零样本学习算法.基于交叉分布对齐变分自编码器网络框架,将当前任务与已学任务的视觉潜层向量对齐,增大不同任务潜层空间的相似性.同时,结合选择性再训练方法,提高当前任务模型对已学任务判别能力.针对不同任务,采用已见类视觉-隐向量和未见类语义-隐向量训练独立的分类器,实现零样本图像分类.在4个标准数据集上的实验表明文中算法能有效实现持续零样本识别任务,缓解算法的灾难性遗忘.     

When Does Communication Learning Need Hierarchical Multi-Agent Deep Reinforcement Learning

Abstract

Multi-agent systems need to communicate to coordinate a shared task. We show that a recurrent neural network (RNN) can learn a communication protocol for coordination, even if the actions to coordinate are performed steps after the communication phase. We show that a separation of tasks with different temporal scale is necessary for successful learning. We contribute a hierarchical deep reinforcement learning model for multi-agent systems that separates the communication and coordination task from the action picking through a hierarchical policy. We further on show, that a separation of concerns in communication is beneficial but not necessary. As a testbed, we propose the Dungeon Lever Game and we extend the Differentiable Inter-Agent Learning (DIAL) framework. We present and compare results from different model variations on the Dungeon Lever Game.     

Haptic User Interfaces for the Visually Impaired: Implications for Haptically Enhanced Science Learning Systems

The overall quality of haptic user interfaces designed to support visually impaired students' science learning through sensorial feedback was systematically studied to investigate task performance and user behavior. Fourteen 6th- to 11th-grade students with visual impairments recruited from a state-funded blind school were asked to perform three main tasks (i.e., menu selection, structure exploration, and force recognition) using haptic user interfaces and a haptic device. This study used several dependent measures that are categorized into three types of variables: (a) task performance including success rate, workload, and task completion time; (b) user behavior defined as cursor movements proportionately represented from the user's cursor positional data; and (c) user preference. Results showed that interface type has significant effects on task performance, user behavior, and user preference, with varying degree of impact to participants with severe visual impairments performing the tasks. The results of this study as well as a set of refined design guidelines and principles should provide insights to the future research of haptic user interfaces that can be used when developing haptically enhanced science learning systems for the visually impaired.     

Group online adaptive learning

Sharing information among multiple learning agents can accelerate learning. It could be particularly useful if learners operate in continuously changing environments, because a learner could benefit from previous experience of another learner to adapt to their new environment. Such group-adaptive learning has numerous applications, from predicting financial time-series, through content recommendation systems, to visual understanding for adaptive autonomous agents. Here we address the problem in the context of online adaptive learning. We formally define the learning settings of Group Online Adaptive Learning and derive an algorithm named Shared Online Adaptive Learning (SOAL) to address it. SOAL avoids explicitly modeling changes or their dynamics, and instead shares information continuously. The key idea is that learners share a common small pool of experts, which they can use in a weighted adaptive way. We define group adaptive regret and prove that SOAL maintains known bounds on the adaptive regret obtained for single adaptive learners. Furthermore, it quickly adapts when learning tasks are related to each other. We demonstrate the benefits of the approach for two domains: vision and text. First, in the visual domain, we study a visual navigation task where a robot learns to navigate based on outdoor video scenes. We show how navigation can improve when knowledge from other robots in related scenes is available. Second, in the text domain, we create a new dataset for the task of assigning submitted papers to relevant editors. This is, inherently, an adaptive learning task due to the dynamic nature of research fields evolving in time. We show how learning to assign editors improves when knowledge from other editors is available. Together, these results demonstrate the benefits for sharing information across learners in concurrently changing environments.

     

用于社交媒体的中文命名实体识别

社交领域的中文命名实体识别(NER)是自然语言处理(NLP)中一项重要的基础任务。目前基于词粒度信息或者外部知识的中文命名实体识别方法,都会受到中文分词(CWS)和溢出词(OOV)等问题的影响。因此,该文提出了一种基于字符的使用位置编码和多种注意力的对抗学习模型。联合使用位置编码和多头注意力能够更好地捕获字序间的依赖关系,而使用空间注意力的判别器则能改善对外部知识的提取效果。该文模型分别在Weibo2015 数据集和Weibo2017数据集上进行了实验,实验结果中的F₁值分别为56.79%和60.62%。与多个基线模型相比,该文提出的模型性能更优。     

脑激活任务区分度的分析及应用研究

大脑在执行不同类型任务时激活模式各不相同,变化很大,各个脑区的变化程度也不同。据此,提出任务区分度计算这一全新的方法。用相似性度量对任务态功能磁共振成像（functional Magnetic Resonance Imaging,fMRI）分析,衡量大脑在执行不同条件时各个脑区激活模式的区分程度,揭示大脑各个区域对任务的表征能力。实验对正常人和狂躁症患者记忆提取任务的fMRI数据进行分析,使用皮尔逊相关分析、余弦相似度分析和欧几里德距离计算3种常用的相似性度量方法,并计算各个脑区的任务区分度。结果表明区分度较高的脑区参与记忆、注意和视觉信息等功能,表明了该方法的准确性和科学性。狂躁症患者在负责记忆和注意等脑区的任务区分度较正常人低,表明患者脑功能受损。此外,研究还发现基于皮尔逊相关分析的区分度计算表现较好。通过与SVM方法的对比证明了该方法在区分不同任务的激活模式时的优越性。综上,基于相似性度量的脑激活任务区分度的方法能够适用于任务态fMRI分析及其相应的脑功能分析。     

深度对抗视觉生成综述

深度视觉生成是计算机视觉领域的热门方向,旨在使计算机能够根据输入数据自动生成预期的视觉内容。深度视觉生成使用人工智能技术赋能相关产业,推动产业自动化、智能化改革与转型。生成对抗网络（generative adversarial networks,GANs）是深度视觉生成的有效工具,近年来受到极大关注,成为快速发展的研究方向。GANs能够接收多种模态的输入数据,包括噪声、图像、文本和视频,以对抗博弈的模式进行图像生成和视频生成,已成功应用于多项视觉生成任务。利用GANs实现真实的、多样化和可控的视觉生成具有重要的研究意义。本文对近年来深度对抗视觉生成的相关工作进行综述。首先介绍深度视觉生成背景及典型生成模型,然后根据深度对抗视觉生成的主流任务概述相关算法,总结深度对抗视觉生成目前面临的痛点问题,在此基础上分析深度对抗视觉生成的未来发展趋势。     

融合对抗学习的因果关系抽取

因果关系抽取在事件预测、情景生成、问答以及文本蕴涵等任务上都有重要的应用价值.但多数现有的因果关系抽取方法都需要人工定义模式和约束,且严重依赖知识库.为此,本文利用生成式对抗网络（Generative adversarial networks,GAN）的对抗学习特性,将带注意力机制的双向门控循环单元神经网络（Bidirectional gated recurrent units networks,BGRU）与对抗学习相融合,通过重定义生成模型和判别模型,基本的因果关系抽取网络能够与判别网络形成对抗,进而从因果关系解释信息中获得高区分度的特征.实验结果表明,与当前用于因果关系抽取的方法相比较,该方法表现出更优的抽取效果.     

UIFGAN: An unsupervised continual-learning generative adversarial network for unified image fusion

In this paper, we propose a novel unsupervised continual-learning generative adversarial network for unified image fusion, termed as UIFGAN. In our model, for multiple image fusion tasks, a generative adversarial network for training a single model with memory in a continual-learning manner is proposed, rather than training an individual model for each fusion task or jointly training multiple tasks. We use elastic weight consolidation to avoid forgetting what has been learned from previous tasks when training multiple tasks sequentially. In each task, the generation of the fused image comes from the adversarial learning between a generator and a discriminator. Meanwhile, a max-gradient loss function is adopted for forcing the fused image to obtain richer texture details of the corresponding regions in two source images, which applies to most typical image fusion tasks. Extensive experiments on multi-exposure, multi-modal and multi-focus image fusion tasks demonstrate the advantages of our method over the state-of-the-art approaches.     

基于对比约束的可解释小样本学习

不同于基于大规模监督的深度学习方法,小样本学习旨在从极少的几个样本中学习这类样本的特性,其更符合人脑的视觉认知机制.近年来,小样本学习受到很多学者关注,他们联合元学习训练模式与度量学习理论,挖掘查询集(无标记样本)和支持集(少量标记样本)在特征空间的语义相似距离,取得不错的小样本分类性能.然而,这些方法的可解释性偏弱,不能为用户提供一种便于直观理解的小样本推理过程.为此,提出一种基于区域注意力机制的小样本分类网络INT-FSL,旨在揭示小样本分类中的2个关键问题：1)图像哪些关键位置的视觉特征在决策中发挥了重要作用;2)这些关键位置的视觉特征能体现哪些类别的特性.除此之外,尝试在每个小样本元任务中设计全局和局部2种对比学习机制,利用数据内部信息来缓解小样本场景中的监督信息匮乏问题.在3个真实图像数据集上进行了详细的实验分析,结果表明：所提方法INT-FSL不仅能有效提升当前小样本学习方法的分类性能,还具备良好的过程可解释性.     

Maintaining filter structure: A Gabor-based convolutional neural network for image analysis

In image segmentation and classification tasks, utilizing filters based on the target object improves performance and requires less training data. We use the Gabor filter as initialization to gain more discriminative power. Considering the mechanism of the error backpropagation procedure to learn the data, after a few updates, filters will lose their initial structure. In this paper, we modify the updating rule in Gradient Descent to maintain the properties of Gabor filters. We use the Left Ventricle (LV) segmentation task and handwritten digit classification task to evaluate our proposed method. We compare Gabor initialization with random initialization and transfer learning initialization using convolutional autoencoders and convolutional networks. We experimented with noisy data and we reduced the amount of training data to compare how different methods of initialization can deal with these matters. The results show that the pixel predictions for the segmentation task are highly correlated with the ground truth. In the classification task, in addition to Gabor and random initialization, we initialized the network using pre-trained weights obtained from a convolutional Autoencoder using two different data sets and pre-trained weights obtained from a convolutional neural network. The experiments confirm the out-performance of Gabor filters comparing to the other initialization method even when using noisy inputs and a lesser amount of training data.     

基于生成对抗网络的红外图像数据增强

深度学习在视觉任务中的良好表现很大程度上依赖于海量的数据和计算力的提升,但是在很多实际项目中通常难以提供足够的数据来完成任务。针对某些情况下红外图像少且难以获得的问题,提出一种基于彩色图像生成红外图像的方法来获取更多的红外图像数据。首先,用现有的彩色图像和红外图像数据构建成对的数据集;然后,基于卷积神经网络、转置卷积神经网络构建生成对抗网络（GAN）模型的生成器和鉴别器;接着,基于成对的数据集来训练GAN模型,直到生成器和鉴别器之间达到纳什平衡状态;最后,用训练好的生成器将彩色图像从彩色域变换到红外域。基于定量评估标准对实验结果进行了评估,结果表明,所提方法可以生成高质量的红外图像,并且相较于在损失函数中不加正则化项,在损失函数中加入L1和L2正则化约束后,该方法的FID分数值平均分别降低了23.95和20.89。作为一种无监督的数据增强方法,该方法也可以被应用于其他缺少数据的目标识别、目标检测、数据不平衡等视觉任务中。     

关联学习:关联关系挖掘新视角

关联关系挖掘与发现是大数据挖掘与分析的重要基础,现有的关联关系挖掘方法多是对数据进行统计分析,对未知数据缺少关联判别作用.尝试从学习的角度进行关联关系挖掘,给出了关联学习的形式化定义和相关概念,并根据关联学习定义构建学习数据集.具体地构建了2类关联图像数据集(two class associated image data sets, TAID),利用卷积神经网络提取关联特征,然后分别用softmax函数和K近邻算法判别关联关系,基于此提出3种关联关系判别器:关联图像卷积神经网络判别器(associated image convolutional neural network discriminator, AICNN)、关联图像LeNet判别器(associated image LeNet discriminator, AILeNet)和关联图像K近邻判别器(associated image K-nearest neighbor discriminator, AIKNN).3种关联判别器在TAID数据集上进行测试,AICNN在64×64像素90 000个训练样本上的判别精度达0.821 7,AILeNet在256×256像素22 500个训练样本上的判别精度达0.845 6,AIKNN在256×256像素22 500个训练样本上的判别精度达到0.866 4.这3种关联判别器有效地证明了学习角度挖掘关联关系的可行性.     

Bayesian computation in recurrent neural circuits

A large number of human psychophysical results have been successfully explained in recent years using Bayesian models. However, the neural implementation of such models remains largely unclear. In this article, we show that a network architecture commonly used to model the cerebral cortex can implement Bayesian inference for an arbitrary hidden Markov model. We illustrate the approach using an orientation discrimination task and a visual motion detection task. In the case of orientation discrimination, we show that the model network can infer the posterior distribution over orientations and correctly estimate stimulus orientation in the presence of significant noise. In the case of motion detection, we show that the resulting model network exhibits direction selectivity and correctly computes the posterior probabilities over motion direction and position. When used to solve the well-known random dots motion discrimination task, the model generates responses that mimic the activities of evidence-accumulating neurons in cortical areas LIP and FEF. The framework we introduce posits a new interpretation of cortical activities in terms of log posterior probabilities of stimuli occurring in the natural world.     

基于强化学习的对抗预训练语言建模方法

在大规模无监督语料上的BERT、XLNet等预训练语言模型,通常采用基于交叉熵损失函数的语言建模任务进行训练。模型的评价标准则采用困惑度或者模型在其他下游自然语言处理任务中的性能指标,存在损失函数和评测指标不匹配等问题。为解决这些问题,该文提出一种结合强化学习的对抗预训练语言模型RL-XLNet(Reinforcement Learning-XLNet)。RL-XLNet采用对抗训练方式训练一个生成器,基于上下文预测选定词,并训练一个判别器判断生成器预测的词是否正确。通过对抗网络生成器和判别器的相互促进作用,强化生成器对语义的理解,提高模型的学习能力。由于在文本生成过程中存在采样过程,导致最终的损失无法直接进行回传,故提出采用强化学习的方式对生成器进行训练。基于通用语言理解评估基准(GLUE Benchmark)和斯坦福问答任务(SQuAD 1.1)的实验,结果表明,与现有BERT、XLNet方法相比,RL-XLNet模型在多项任务中的性能上表现出较明显的优势: 在GLUE的六个任务中排名第1,一个任务排名第2,一个任务排名第3。在SQuAD 1.1任务中F₁值排名第1。考虑到运算资源有限,基于小语料集的模型性能也达到了领域先进水平。     

On hierarchical modelling of motion for workflow analysis from overhead view

Understanding human behaviour is a high level perceptual problem, one which is often dominated by the contextual knowledge of the environment, and where concerns such as occlusion, scene clutter and high within-class variations are commonplace. Nonetheless, such understanding is highly desirable for automated visual surveillance. We consider this problem in a context of a workflow analysis within an industrial environment. The hierarchical nature of the workflow is exploited to split the problem into ‘activity’ and ‘task’ recognition. In this, sequences of low level activities are examined for instances of a task while the remainder are labelled as background. An initial prediction of activity is obtained using shape and motion based features of the moving blob of interest. A sequence of these activities is further adjusted by a probabilistic analysis of transitions between activities using hidden Markov models (HMMs). In task detection, HMMs are arranged to handle the activities within each task. Two separate HMMs for task and background compete for an incoming sequence of activities. Imagery derived from a camera mounted overhead the target scene has been chosen over the more conventional oblique views (from the side) as this view does not suffer from as much occlusion, and it poses a manageable detection and tracking problem while still retaining powerful cues as to the workflow patterns. We evaluate our approach both in activity and task detection on a challenging dataset of surveillance of human operators in a car manufacturing plant. The experimental results show that our hierarchical approach can automatically segment the timeline and spatially localize a series of predefined tasks that are performed to complete a workflow.     

基于强化学习的移动边缘计算任务卸载方法

最佳卸载策略直接影响移动计算任务卸载的时延与能耗,因此提出基于强化学习方法的移动边缘计算任务卸载方法。首先对移动设备的计算任务卸载形式展开具体分析,并基于分析结果获取计算任务卸载能量消耗、发射功率、传输速率等相关参数值,以此建立移动边缘计算任务卸载模型。最后基于建立的卸载模型结合Q-Learning算法对计算任务实施强化学习,找出计算任务的最佳卸载策略,从而实现移动边缘计算任务的实时卸载。实验结果表明,使用强化学习方法开展移动边缘计算任务卸载时,卸载能耗低、时延小。     

Phasic norepinephrine: a neural interrupt signal for unexpected events

Extensive animal studies indicate that the neuromodulator norepinephrine plays an important role in specific aspects of vigilance, attention and learning, putatively serving as a neural interrupt or reset function. The activity of norepinephrine-releasing neurons in the locus coeruleus during attentional tasks is modulated not only by the animal's level of engagement and the sensory inputs, but also by temporally rich aspects of internal decision-making processes. Here, we propose that it is unexpected changes in the world within the context of a task that activate the noradrenergic interrupt signal. We quantify this idea in a Bayesian model of a well-studied visual discrimination task, demonstrating that the model captures a rich repertoire of noradrenergic responses at the sub-second temporal resolution.