Conversational transfer learning for emotion recognition

Recognizing emotions in conversations is a challenging task due to the presence of contextual dependencies governed by self- and inter-personal influences. Recent approaches have focused on modeling these dependencies primarily via supervised learning. However, purely supervised strategies demand large amounts of annotated data, which is lacking in most of the available corpora in this task. To tackle this challenge, we look at transfer learning approaches as a viable alternative. Given the large amount of available conversational data, we investigate whether generative conversational models can be leveraged to transfer affective knowledge for detecting emotions in context. We propose an approach, TL-ERC, where we pre-train a hierarchical dialogue model on multi-turn conversations (source) and then transfer its parameters to a conversational emotion classifier (target). In addition to the popular practice of using pre-trained sentence encoders, our approach also incorporates recurrent parameters that model inter-sentential context across the whole conversation. Based on this idea, we perform several experiments across multiple datasets and find improvement in performance and robustness against limited training data. TL-ERC also achieves better validation performances in significantly fewer epochs. Overall, we infer that knowledge acquired from dialogue generators can indeed help recognize emotions in conversations.     

Semi-supervised labeling: a proposed methodology for labeling the twitter datasets

Twitter has nowadays become a trending microblogging and social media platform for news and discussions. Since the dramatic increase in its platform has additionally set off a dramatic increase in spam utilization in this platform. For Supervised machine learning, one always finds a need to have a labeled dataset of Twitter. It is desirable to design a semi-supervised labeling technique for labeling newly prepared recent datasets. To prepare the labeled dataset lot of human affords are required. This issue has motivated us to propose an efficient approach for preparing labeled datasets so that time can be saved and human errors can be avoided. Our proposed approach relies on readily available features in real-time for better performance and wider applicability. This work aims at collecting the most recent tweets of a user using Twitter streaming and prepare a recent dataset of Twitter. Finally, a semi-supervised machine learning algorithm based on the self-training technique was designed for labeling the tweets. Semi-supervised support vector machine and semi-supervised decision tree classifiers were used as base classifiers in the self-training technique. Further, the authors have applied K means clustering algorithm to the tweets based on the tweet content. The principled novel approach is an ensemble of semi-supervised and unsupervised learning wherein it was found that semi-supervised algorithms are more accurate in prediction than unsupervised ones. To effectively assign the labels to the tweets, authors have implemented the concept of voting in this novel approach and the label pre-directed by the majority voting classifier is the actual label assigned to the tweet dataset. Maximum accuracy of 99.0% has been reported in this paper using a majority voting classifier for spam labeling.

     

Algebraic manipulation of scientific datasets

We investigate algebraic processing strategies for large numeric datasets equipped with a (possibly irregular) grid structure. Such datasets arise, for example, in computational simulations, observation networks, medical imaging, and 2-D and 3-D rendering. Existing approaches for manipulating these datasets are incomplete: The performance of SQL queries for manipulating large numeric datasets is not competitive with specialized tools. Database extensions for processing multidimensional discrete data can only model regular, rectilinear grids. Visualization software libraries are designed to process arbitrary gridded datasets efficiently, but no algebra has been developed to simplify their use and afford optimization. Further, these libraries are data dependent – physical changes to data representation or organization break user programs. In this paper, we present an algebra of gridfields for manipulating arbitrary gridded datasets, algebraic optimization techniques, and an implementation backed by experimental results. We compare our techniques to those of Geographic Information Systems (GIS) and visualization software libraries, using real examples from an Environmental Observation and Forecasting System. We find that our approach can express optimized plans inaccessible to other techniques, resulting in improved performance with reduced programming effort.     

A novel pruning approach using expert knowledge for data-specific pruning

Classification is an important data mining task that discovers hidden knowledge from the labeled datasets. Most approaches to pruning assume that all dataset are equally uniform and equally important, so they apply equal pruning to all the datasets. However, in real-world classification problems, all the datasets are not equal and considering equal pruning rate during pruning tends to generate a decision tree with large size and high misclassification rate. We approach the problem by first investigating the properties of each dataset and then deriving data-specific pruning value using expert knowledge which is used to design pruning techniques to prune decision trees close to perfection. An efficient pruning algorithm dubbed EKBP is proposed and is very general as we are free to use any learning algorithm as the base classifier. We have implemented our proposed solution and experimentally verified its effectiveness with forty real world benchmark dataset from UCI machine learning repository. In all these experiments, the proposed approach shows it can dramatically reduce the tree size while enhancing or retaining the level of accuracy.     

Transfer learning between crop types for semantic segmentation of crops versus weeds in precision agriculture

Agricultural robots rely on semantic segmentation for distinguishing between crops and weeds to perform selective treatments and increase yield and crop health while reducing the amount of chemicals used. Deep‐learning approaches have recently achieved both excellent classification performance and real‐time execution. However, these techniques also rely on a large amount of training data, requiring a substantial labeling effort, both of which are scarce in precision agriculture. Additional design efforts are required to achieve commercially viable performance levels under varying environmental conditions and crop growth stages. In this paper, we explore the role of knowledge transfer between deep‐learning‐based classifiers for different crop types, with the goal of reducing the retraining time and labeling efforts required for a new crop. We examine the classification performance on three datasets with different crop types and containing a variety of weeds and compare the performance and retraining efforts required when using data labeled at pixel level with partially labeled data obtained through a less time‐consuming procedure of annotating the segmentation output. We show that transfer learning between different crop types is possible and reduces training times for up to 80%. Furthermore, we show that even when the data used for retraining are imperfectly annotated, the classification performance is within 2% of that of networks trained with laboriously annotated pixel‐precision data.     

基于伪标签深度学习的高光谱影像半监督分类

研究了一种仅利用少量标记点训练深度卷积神经网络并对高光谱影像进行分类的方法。以图像分割获得的同质区增加训练样本数目；借助这些增加的样本训练初始分类器并预测所有未知点的初始类别；将每一初始类别聚集为适当的类簇，以类簇号作为伪标签对深度卷积网进行预训练；最后利用经过同质区增加的训练样本精调预训练深度卷积网。实验结果证明新方法可以在仅用少量实际标记样本的情况下成功地训练深度卷积网，对高光谱数据进行有效分类。     

Improving robustness of industrial object detection by automatic generation of synthetic images from CAD models

Object detection (OD) is used for visual quality control in factories. Images that compose training datasets are often collected directly from the production line and labeled with bounding boxes manually. Such data represent well the inference context but might lack diversity, implying a risk of overfitting. To address this issue, we propose a dataset construction method based on an automated pipeline, which receives a CAD model of an object and returns a set of realistic synthetic labeled images (code publicly available). Our approach can be easily used by non-expert users and is relevant for industrial applications, where CAD models are widely available. We performed experiments to compare the use of datasets obtained by the two different ways—collecting and labeling real images or applying the proposed automated pipeline—in the classification of five different industrial parts. To ensure that both approaches can be used without deep learning expertise, all training parameters were kept fixed during these experiments. In our results, both methods were successful for some objects but failed for others. However, we have shown that the combined use of real and synthetic images led to better results. This finding has the potential to make industrial OD models more robust to poor data collection and labeling errors, without increasing the difficulty of the training process.     

Semi-supervised classification and betweenness computation on large, sparse, directed graphs

This work addresses graph-based semi-supervised classification and betweenness computation in large, sparse, networks (several millions of nodes). The objective of semi-supervised classification is to assign a label to unlabeled nodes using the whole topology of the graph and the labeling at our disposal. Two approaches are developed to avoid explicit computation of pairwise proximity between the nodes of the graph, which would be impractical for graphs containing millions of nodes. The first approach directly computes, for each class, the sum of the similarities between the nodes to classify and the labeled nodes of the class, as suggested initially in [1] and [2]. Along this approach, two algorithms exploiting different state-of-the-art kernels on a graph are developed. The same strategy can also be used in order to compute a betweenness measure. The second approach works on a trellis structure built from biased random walks on the graph, extending an idea introduced in [3]. These random walks allow to define a biased bounded betweenness for the nodes of interest, defined separately for each class. All the proposed algorithms have a linear computing time in the number of edges while providing good results, and hence are applicable to large sparse networks. They are empirically validated on medium-size standard data sets and are shown to be competitive with state-of-the-art techniques. Finally, we processed a novel data set, which is made available for benchmarking, for multi-class classification in a large network: the U.S. patents citation network containing 3M nodes (of six different classes) and 38M edges. The three proposed algorithms achieve competitive results (around 85% classification rate) on this large network-they classify the unlabeled nodes within a few minutes on a standard workstation.     

CoralSeg: Learning coral segmentation from sparse annotations

Robotic advances and developments in sensors and acquisition systems facilitate the collection of survey data in remote and challenging scenarios. Semantic segmentation, which attempts to provide per‐pixel semantic labels, is an essential task when processing such data. Recent advances in deep learning approaches have boosted this task's performance. Unfortunately, these methods need large amounts of labeled data, which is usually a challenge in many domains. In many environmental monitoring instances, such as the coral reef example studied here, data labeling demands expert knowledge and is costly. Therefore, many data sets often present scarce and sparse image annotations or remain untouched in image libraries. This study proposes and validates an effective approach for learning semantic segmentation models from sparsely labeled data. Based on augmenting sparse annotations with the proposed adaptive superpixel segmentation propagation, we obtain similar results as if training with dense annotations, significantly reducing the labeling effort. We perform an in‐depth analysis of our labeling augmentation method as well as of different neural network architectures and loss functions for semantic segmentation. We demonstrate the effectiveness of our approach on publicly available data sets of different real domains, with the emphasis on underwater scenarios—specifically, coral reef semantic segmentation. We release new labeled data as well as an encoder trained on half a million coral reef images, which is shown to facilitate the generalization to new coral scenarios.     

Multi-context unsupervised domain adaption for HEp-2 cell classification using maximum partial classifier discrepancy

Identifying human epithelial-2 (HEp-2) cells in indirect immune fluorescence (IIF) is the most commonly used method for the diagnosis of autoimmune diseases. Although supervised deep learning networks have made remarkable progress on HEp-2 cell staining pattern classification, the high-performance relies on a large amount of labeled training data. Unfortunately, large-scale labeled datasets are scarce due to the expensive costs of labeling medical images. Therefore, we propose an unsupervised domain adaption (UDA) model, namely MDA-MPCD, to classify unlabeled HEp-2 cell samples. The proposed model involves two major aspects: (a) multi-context domain adaption (MDA) generator and (b) maximum partial classifier discrepancy (MPCD) architecture. The MDA generator can extract multi-context features from complex cell images while providing more comprehensive and diverse information for the classifier. The MPCD architecture, utilizing the mapping variation of feature transfer, focuses on the discrepancy from the cross-domain gap by separating the activations in the classifier. The proposed model dominates all comparison methods during evaluation, achieving 85.35% and 96.08% mean accuracy on two UDA tasks, respectively. Furthermore, the model is demonstrated to adapt from rich label domain to unlabeled domain by detailed ablation experiments and visualization results. The proposed MDA-MPCD has potential as a clinical scheme, enabling effective and accurate classification of HEp-2 cell staining pattern without labeling the target data.

     

Solution path for manifold regularized semisupervised classification

Traditional learning algorithms use only labeled data for training. However, labeled examples are often difficult or time consuming to obtain since they require substantial human labeling efforts. On the other hand, unlabeled data are often relatively easy to collect. Semisupervised learning addresses this problem by using large quantities of unlabeled data with labeled data to build better learning algorithms. In this paper, we use the manifold regularization approach to formulate the semisupervised learning problem where a regularization framework which balances a tradeoff between loss and penalty is established. We investigate different implementations of the loss function and identify the methods which have the least computational expense. The regularization hyperparameter, which determines the balance between loss and penalty, is crucial to model selection. Accordingly, we derive an algorithm that can fit the entire path of solutions for every value of the hyperparameter. Its computational complexity after preprocessing is quadratic only in the number of labeled examples rather than the total number of labeled and unlabeled examples.     

On the reliable detection of concept drift from streaming unlabeled data

Classifiers deployed in the real world operate in a dynamic environment, where the data distribution can change over time. These changes, referred to as concept drift, can cause the predictive performance of the classifier to drop over time, thereby making it obsolete. To be of any real use, these classifiers need to detect drifts and be able to adapt to them, over time. Detecting drifts has traditionally been approached as a supervised task, with labeled data constantly being used for validating the learned model. Although effective in detecting drifts, these techniques are impractical, as labeling is a difficult, costly and time consuming activity. On the other hand, unsupervised change detection techniques are unreliable, as they produce a large number of false alarms. The inefficacy of the unsupervised techniques stems from the exclusion of the characteristics of the learned classifier, from the detection process. In this paper, we propose the Margin Density Drift Detection (MD3) algorithm, which tracks the number of samples in the uncertainty region of a classifier, as a metric to detect drift. The MD3 algorithm is a distribution independent, application independent, model independent, unsupervised and incremental algorithm for reliably detecting drifts from data streams. Experimental evaluation on 6 drift induced datasets and 4 additional datasets from the cybersecurity domain demonstrates that the MD3 approach can reliably detect drifts, with significantly fewer false alarms compared to unsupervised feature based drift detectors. At the same time, it produces performance comparable to that of a fully labeled drift detector. The reduced false alarms enables the signaling of drifts only when they are most likely to affect classification performance. As such, the MD3 approach leads to a detection scheme which is credible, label efficient and general in its applicability.     

Multi-label active learning by model guided distribution matching

Multi-label learning is an effective framework for learning with objects that have multiple semantic labels, and has been successfully applied into many real-world tasks. In contrast with traditional single-label learning, the cost of labeling a multi-label example is rather high, thus it becomes an important task to train an effectivemulti-label learning model with as few labeled examples as possible. Active learning, which actively selects the most valuable data to query their labels, is the most important approach to reduce labeling cost. In this paper, we propose a novel approach MADM for batch mode multi-label active learning. On one hand, MADM exploits representativeness and diversity in both the feature and label space by matching the distribution between labeled and unlabeled data. On the other hand, it tends to query predicted positive instances, which are expected to be more informative than negative ones. Experiments on benchmark datasets demonstrate that the proposed approach can reduce the labeling cost significantly.     

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。     

Random linear interpolation data augmentation for person re-identification

Person Re-Identification (person re-ID) is an image retrieval task which identifies the same person in different camera views. Generally, a good person re-ID model requires a large dataset containing over 100000 images to reduce the risk of over-fitting. Most current handcrafted person re-ID datasets, however, are insufficient for training a learning model with high generalization ability. In addition, the lacking of images with various levels of occlusion is still remaining in most existing datasets. Motivated by these two problems, this paper proposes a new data augmentation method called Random Linear Interpolation that can enlarge the sizes of person re-ID datasets and improve the generalization ability of the learning model. The key enabler of our approach is generating fused images by interpolating pairs of original images. In other words, the innovation of the proposed approach is considering data augmentation between two random samples. Plenty of experimental results demonstrates that the proposed method is effective to improve baseline models. On Market1501 and DukeMTMC-reID datasets, our approach can achieve 92.71% and 82.19% rank-1 accuracy, respectively.

     

Incremental predictive clustering trees for online semi-supervised multi-target regression

In many application settings, labeling data examples is a costly endeavor, while unlabeled examples are abundant and cheap to produce. Labeling examples can be particularly problematic in an online setting, where there can be arbitrarily many examples that arrive at high frequencies. It is also problematic when we need to predict complex values (e.g., multiple real values), a task that has started receiving considerable attention, but mostly in the batch setting. In this paper, we propose a method for online semi-supervised multi-target regression. It is based on incremental trees for multi-target regression and the predictive clustering framework. Furthermore, it utilizes unlabeled examples to improve its predictive performance as compared to using just the labeled examples. We compare the proposed iSOUP-PCT method with supervised tree methods, which do not use unlabeled examples, and to an oracle method, which uses unlabeled examples as though they were labeled. Additionally, we compare the proposed method to the available state-of-the-art methods. The method achieves good predictive performance on account of increased consumption of computational resources as compared to its supervised variant. The proposed method also beats the state-of-the-art in the case of very few labeled examples in terms of performance, while achieving comparable performance when the labeled examples are more common.

     

Entity attribute discovery and clustering from online reviews

The rapid increase of user-generated content (UGC) is a rich source for reputation management of entities, products, and services. Looking at online product reviews as a concrete example, in reviews, customers usually give opinions on multiple attributes of products, therefore the challenge is to automatically extract and cluster attributes that are mentioned. In this paper, we investigate efficient attribute extraction models using a semi-supervised approach. Specifically, we formulate the attribute extraction issue as a sequence labeling task and design a bootstrapped schema to train the extraction models by leveraging a small quantity of labeled reviews and a larger number of unlabeled reviews. In addition, we propose a clustering By committee (CBC) approach to cluster attributes according to their semantic similarity. Experimental results on real world datasets show that the proposed approach is effective.     

Feature matching and instance reweighting with transfer learning for human activity recognition using smartphone

Human activity recognition using smartphone has been attracting great interest. Since collecting large amount of labeled data is expensive and time-consuming for conventional machine learning techniques, transfer learning techniques have been proposed for activity recognition. However, existing transfer learning techniques typically rely on feature matching based on global domain shift and lack considering the intra-class knowledge transfer. In this paper, a novel transfer learning technique is proposed for cross-domain activity recognition, which can properly integrate feature matching and instance reweighting across the source and target domain in principled dimensionality reduction. The experiments using three real datasets demonstrate that the proposed scheme can achieve much higher precision (92%), recall (91%), and F1-score (92%), in comparison with the existing schemes.

     

Automatic generation of water distribution systems based on GIS data

In the field of water distribution system (WDS) analysis, case study research is needed for testing or benchmarking optimisation strategies and newly developed software. However, data availability for the investigation of real cases is limited due to time and cost needed for data collection and model setup. We present a new algorithm that addresses this problem by generating WDSs from GIS using population density, housing density and elevation as input data. We show that the resulting WDSs are comparable to actual systems in terms of network properties and hydraulic performance. For example, comparing the pressure heads for an actual and a generated WDS results in pressure head differences of ±4 m or less for 75% of the supply area. Although elements like valves and pumps are not included, the new methodology can provide water distribution systems of varying levels of complexity (e.g., network layouts, connectivity, etc.) to allow testing design/optimisation algorithms on a large number of networks. The new approach can be used to estimate the construction costs of planned WDSs aimed at addressing population growth or at comparisons of different expansion strategies in growth corridors.