Multi-view Opinion Mining with Deep Learning

Size: px
Start display at page:

Download "Multi-view Opinion Mining with Deep Learning"

Transcription

1 Neural Processing Letters manuscript No. (will be inserted by the editor) Multi-view Opinion Mining with Deep Learning Ping Huang Xijiong Xie Shiliang Sun Received: date / Accepted: date Abstract With the explosive growth of social media on the Internet, people are expressing an increasing number of opinions. As for objectives like business decision making and public opinion analysis, how to make the best of these precious opinionated words is a new challenge in the field of NLP. The field of opinion mining, or sentiment analysis, has become active in recent years. Since different kinds of deep neural networks differ in their structures, they are probably extracting different features. We investigated whether features generated by heterogeneous deep neural networks can be combined by multiview learning to improve the overall performance. With document level opinion mining being the objective, we implemented multi-view learning based on heterogeneous deep neural networks. Experiments show that multi-view learning utilizing these heterogeneous features outperforms single-view deep neural networks. Our framework makes better use of single-view data. Keywords Multi-view learning Opinion mining Deep learning Heterogeneous neural networks Ping Huang Department of Computer Science and Technology, East China Normal University, 3663 North Zhongshan Road, Shanghai , P.R. China phuang95@outlook.com Xijiong Xie The School of Information Science and Engineering, Ningbo University, Zhejiang , China xjxie11@gmail.com Shiliang Sun Department of Computer Science and Technology, East China Normal University, 3663 North Zhongshan Road, Shanghai , P.R. China (Corresponding Author) slsun@cs.ecnu.edu.cn Ping Huang and Xijiong Xie contributed equally to this work

2 2 Ping Huang et al. 1 Introduction With the explosive growth of Internet social media, people are expressing an increasing number of opinions. The field of opinion mining, which exploits these precious opinionated words, has accordingly become active in recent years. Typical tasks in this field aims to find out subjective contents in text and tell the sentiment polarities of the speakers. According to the definition in [1], an opinion can be defined as a quintuple (e i, a ij, s ijkl, h k, t l ), where e i stands for the entity in the opinion, a ij stands for the aspect of the entity, h k stands for the opinion holder, t l stands for the time when the opinion was given and s ijkl stands for the opinion contents and opinion holder s sentiment polarity. For example, in the review The screen of this mobile phone is good, the screen is an aspect of the entity mobile phone and a positive sentiment is expressed. With this definition given, the task of opinion mining can be then defined as determining the quintuple, or at least part of it. Tasks in the opinion mining field can be categorized into three levels: document level, sentence level and fine-grained level [2]. Document level opinion mining tasks determine an overall sentiment polarity for a whole document. Sentence level tasks are similar but they determine the sentiment polarities of sentences. Fine-grained opinion mining tasks, however, could handle more detailed information, including entities, aspects, etc. Our research is conducted on document-level. Deep learning as a popular field in machine learning, has also been used in opinion mining tasks. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are typical models used in these tasks. In 2014, Kalchbrenner et al. [3] described a convolutional architecture called dynamic convolutional neural networks (DCNNs) using dynamic k-max over linear sequences. Irsoy and Cardie [4] proposed deep recurrent neural networks (DRNNs) in They are constructed by stacking several layers of RNNs together. Each layer takes outputs from previous layer as its inputs and has its own temporal information representations. Considering that deep neural networks (DNNs) have the ability to automatically extract features from data [5], we are able to extract from the networks the intermediate representations suitable to be used as features. Furthermore, due to the structural differences of kinds of deep neural networks, it is possible that they are using different kinds of intermediate representations. Thus we try combining the deep features from heterogeneous DNNs to see whether they are complementary and whether combining them with multi-view learning methods improves the overall performance. Basing on these thoughts, we propose a new opinion mining framework to leverage heterogeneous DNNs in multi-view learning. The framework takes single-view features as inputs and processes them with heterogeneous DNNs to automatically extract multi-view features for multi-view learning methods. This framework applies multi-view methods on single-view data and attempts to exploit their potentialities.

3 Multi-view Opinion Mining with Deep Learning 3 We implemented the heterogeneous DNN feature based multi-view learning framework to handle document-level opinion mining tasks. First, abstract representations of words are learned from a massive amount of text. Then they are used by heterogeneous DNNs to learn abstract representations of documents. The sets of representations are regarded as distinct views for each document. At last, a multi-view classifier processes the multi-view representations to tell the sentiment polarities of the documents and therefore completes the opinion mining task. The remainder of the paper is organized as follows. In section 2, we introduce multi-view learning methods and how they were used in other works. In section 3, we explain how to choose intermediate representations from deep neural networks and our reasons basing on the architectures of these neural networks. In section 4, we describe our model which is based on classical DNNs from beginning to end. In section 5, the model is tested on a small-scale data set. In section 6, we conclude our work briefly. 2 Multi-view learning in other works Multi-view data are quite common in real world applications. Describing real world entities by only one measuring method can hardly depict all sides of them, while multi-view data generated by multiple measuring methods cover more details. In pictures, multi-view data can be from pixel level and texture information. For videos, all the frames and the audio can be two view of data. Multi-view learning is an emerging field in machine learning, which does research on exploiting multi-view data for a single object to improve overall performance. Recently, multi-view learning has been widely applied in numerous fields such as image retrieval and ranking [6 8], human pose recovery [9, 10] and so on. As a method combining multiple data sets, it can also be seen as data fusion. Due to different usage of multi-view data, a high level categorization [11] divides multi-view learning methods to 3 styles: co-training style algorithms, co-regularization style algorithms and margin consistency style algorithms. Co-training methods are basically frameworks based on single-view machine learning methods and are mainly used in semi-supervised learning. For each iteration, unlabeled data on one view can be labeled with the help of classifiers prediction in other views. Co-regularization methods put a regularization term to their objective functions in order to improve consensus among classifiers in the multiple views. Margin consistency methods involve maximum entropy discrimination (MED) and the margin refers to that in the MED classifiers. These methods are based on an assumption that the margins from the views are consistent, which means that the classification confidences from different views are identical [12]. Most multi-view learning methods seek a low-dimensional common subspace to represent multi-view data, then adopt the simple-view algorithm for learning [13 15]. A typical algorithm is canonical correlation analysis (CCA)[15].

4 4 Ping Huang et al. However our method uses multi-view classifiers to combine heterogeneous representations extracted from single-view data. Multi-view learning methods were also seen in two other scenarios in opinion mining. However, their motivation and starting points are completely different from our method. One of them applies multi-view learning to social media that consist of text, pictures and videos and involve users social activities. Niu et al. [16] gave a detailed introduction to methods in this scenario. Tang et al. [17] proposed a multi-view framework to exploit relations among views to help each other select relevant features for social media. The other scenario exploits co-training multi-view methods in cross-lingual problems. Wan [18] used co-training methods and machine learning to make use of unlabeled data in another language. Hajmohammadi et al. [19] sought human s assistance to label a selection of unlabeled corpus in order to find a balance between human cost and performance. 3 Feature extraction from deep neural network In the following paragraphs, separate introductions to two typical DNNs and our thoughts about feature extraction are given. 3.1 Convolutional neural network Starting with several works in 1990s [20, 21], convolutional neural network is one of the earliest types of DNN. They are well designed to discover suitable representations for the raw inputs, which is the key attribute of DNN. A CNN usually consists of several convolutional layers, pooling layers and several fully connected layers at the end. Convolutional layers take a group of adjacent elements (e.g., adjacent pixels in a picture or adjacent words in a sentence) as inputs and compare them to a series of filters to find out valuable patterns [20]. In pooling layers, max pooling, a typical pooling method, finds maximums from local patches of previous outputs. This manipulation matches the most significant patterns in local patches because those less matched patterns have smaller outputs after convolutional layers. Futhermore, pooling layers improve overall tolerance to shifts and distortions [5]. The fully connected layers at the end work as traditional neural networks and leverages features provided by previous layers to complete given tasks. It is worth pointing out that convolutional layers and pooling layers together reduce the dimension of data and extract intermediate representations of data required by successive computations. In conventional machine learning methods before DNNs, either features are carefully designed or raw features that are less meaningful and less effective (e.g., pixel level data for pictures) are directly used [22, 23]. In DNNs, however, representations are automatically learned to be suitable for the task during the training of the whole neural network. This attribute is especially notable in CNNs since their structures are designed for hierarchical feature extraction [5].

5 Multi-view Opinion Mining with Deep Learning 5 Fig. 1 The structure of a typical RNN, and its temporally unfolded version. [5] CNN deep features. In conventional classification approaches, features are directly fed to the neural networks, implying the only job of the neural network is classification. Since fully connected layers in the CNNs act like conventional feed-forward neural networks, we believe intermediate variables between conventional or pooling layers and fully connected layers can be considered as a CNN extraction of the input data. 3.2 Recurrent neural network Recurrent neural networks, another type of DNN, take sequential data as inputs. Given a data point in the sequence, RNNs give an output basing on not only the inputted data point, but also previous data points. In bidirectional RNNs [24], outputs even depend on data both before and after. An instance of RNN uses the same structure, or the cell, to calculate outputs for each data point in the input sequence. RNNs usually do not have circuits within a single time step, which is like conventional feed-forward neural networks, but each calculation process passes information between each other. Sequential data are not necessarily temporally sequential. For instance, gene sequence data are one kind of non-temporal sequence. However, a majority of sequence data are temporal sequence, so the concept of time is usually introduced to help describe sequence data [25]. An input sequence can be defined as (x 1, x 2,..., x T ), where T is the length of the sequence, each data point x t is a real-valued vector. Correspondingly, if the output of an RNN is sequential, it can be represented by (o 1, o 2,..., o T ) in a similar fashion. Many kinds of cells also pass intermediate status that differs from its output. The status at time step t is s t. The structure of a typical RNN is shown in Fig. 1. At time step t, the cell processes not only current data x t, but also previous intermediate status s t 1. For this time step, the cell needs to calculate an output o t and an intermediate status s t which are used in the next time step.

6 6 Ping Huang et al. One of the widely used cells is long short-term memory (LSTM) proposed in 1997 [26]. LSTM cells use the input x t to decide whether a part of the previous status s t 1 is kept or disposed. In 2014, a simplified cell with similar functions was proposed in [27], and is later called gated recurrent unit (GRU). The simplified structure shortens training time, but acts like LSTM when deciding how status information is handled. One apparent difference of GRU from LSTM is that GRU does not distinct output and status, which means its only output serves as the status to be passed to next time step as well. Both of them learn a method to make choices about status information, preventing it from propagating so far that gradients explode or vanish [25] RNN deep features. Information in RNNs travels from time step to time step. RNNs based on LSTM cells or GRU cells are believed to create a memorizing feature to pass information between steps. We speculate the status information passed between steps carries important features of the input sequence that are different from CNNs. 4 Our model Our model consists of four parts. First, data are processed for a better use in stages afterwards. Next, the CNN part and the RNN part are parallel and act as typical DNNs in opinion mining tasks. Chosen intermediate representations are extracted from both DNNs and passed to the last stage. Multi-view classifiers at the end finish the opinion mining task. The overall flowcat of our model can been seen in Fig Input processing Documents are written in natural languages, and there are some detailed problems when presenting text with real-valued data. In most natural language processing tasks, using word vectors instead of conventional one-hot features can directly improve the performance [28]. We use word2vec 1 to convert words into D-dimensional word vectors. Each document consisting of words is therefore converted to a sequence of word vectors. After word2vec finishes training, results are saved as a lookup table, and each word in its vocabulary has a corresponding vector. Documents may contain rare words that are not in the vocabulary. These words are simply removed from the document. Besides, all documents are adjusted to have the same length L so that data processing would be easier. Then all documents can be represented by a 2-dimensional L D matrix and this is the single-view data for a document. The structure is shown in Fig

7 Multi-view Opinion Mining with Deep Learning 7 Fig. 2 The overall flow chart. Fig. 3 The structure of the word2vec part. The single-view data are processed by a CNN and an RNN respectively. They resemble these neural networks in conventional single-view opinion mining tasks, and also output a value to predict the sentiment polarity. 4.2 The CNN part The CNN used in our model is basically the one in [29]. In the CNN, calculation are done on several channels with different window size. Within each channel, data are cut into patches and then processed by a convolutional layer, a ReLU activation layer and a pooling layer. Each channel has its own window size W. Within the convolutional layers, F trainable filters sized W D are compared with the window of word vectors by summing pointwise production. These F values reflect to what extent the data match the filters. Then they are processed by the ReLU activation function. Within the pooling layer, max-pooling finds the maximum per filter, describing how likely the pattern of this filter is seen in this document. Each channel outputs a F -dimensional vector. The results of the channels are concatenated and sent to a perceptron. The cross entropy between the output of the activation function in the perceptron and the observation is used as the loss function. The structure is shown in Fig 4. The intermediate variables extracted to be one of the views are the concatenated outputs of the channels. The reason is twofold. First, the most important structures in the CNN have been finished, so the results here cover the most characteristic operations in the CNN, i.e. convolution and pooling.

8 8 Ping Huang et al. Fig. 4 The structure of the CNN part. Fig. 5 The structure of the RNN part. Second, calculation afterwards is linear classification by a perceptron, while linear classification is one of the most basic tasks in machine learning. If a perceptron handles this task well, it is likely that the intermediate variables at this stage are suitable for classification. 4.3 The RNN part The RNN used in the model adopts an attention mechanism [30] and the cell within is GRU. At each time step, RNN takes in one word vector of the document. In every time step, the GRU cells base their outputs on the current word vector as well as previous status from the previous time step. After all time steps finish calculation, their outputs are collected by the attention layer. It calculates weights for every time step, and outputs a weighted mean. The length of the output vector is configurable. Following parts are similar to the CNN s. A perceptron handles the classification task, and the loss function is the cross entropy. For a similar reason, the deep features extracted from RNN are those variables outputted by the attention layer. The structure is shown in Fig Multi-view classification The two sets of intermediate variables are treated as two views. Given two-view data of documents and observation of sentiment polarities of the documents, a multi-view classifier (e.g., SVM-2K or MVMED) can be trained to fit them. Among the supervised multi-view classifiers based on support vector machine (SVM), one is SVM-2K [31]. It trains SVMs on both views and regularize

9 Multi-view Opinion Mining with Deep Learning 9 them with a constraint of similarity with an ɛ-insensitive term: w Aφ A (x i ) + b A w Bφ B (x i ) b B η i + ɛ (1) where w A, b A (w B, b B ) are the weight and threshold of the first (second) SVM, φ A and φ B are the two feature functions, x is the input and η i is the slack variable. The SVM-2K the following optimization for classifier parameters w A, b A, (w B, b B ) 1 min w A,w B 2 w A w B 2 + c 1 + D n i=1 η i n q Ai + c 2 i=1 n q Bi i=1 s.t. w Aφ A (x i ) + b A w Bφ B (x i ) b B η i + ɛ, y i (w Aφ A (x i ) + b A ) 1 q Ai, y i (w Aφ A (x i )e + b B ) 1 q Bi, q Ai 0, q Bi 0, η i 0, all for (1 i n), where D, c 1, c 2, ɛ are nonnegative parameters and q Ai, q Bi are slack vectors. Denote wa, w B, b A, b B as the solution to this optimization problem. The SVM-2K decision function is then f(x) = 1 2 (w A φ A(x)+b A +w B φ B(x)+b B ). The dual problem of the above optimization problem is given as 1 n (ξi A ξj A K A (x i, x j ) + ξi B ξj B K B (x i, x j )) 2 max ξ A i,ξb i,αa i,αb i + i,j=1 n (αi A + αi B ) i=1 s.t. ξ A i = α A i y i β + i + β i, ξi B = αi B y i + β + i β i, n n ξi A = ξi B = 0, i=1 i=1 0 β + i, β i, β+ i + β i D, 0 α A/B i c 1/2, where αi A, αb i, β+ i, β i are the vectors of Lagrange multipliers. Here we take ɛ = 0. The prediction function for each view is given by n f A/B (x) = ξ A/B i K A/B (x i, x) + b A/B. (4) i=1 (2) (3)

10 10 Ping Huang et al. Multi-view maximum entropy discrimination [12] (MVMED) applies MED to two-view problems by assuming consistent margins from the two views for the fisrt time, which decreases the difficulty of optimization. Some variations [32, 33] are also proposed since MVMED, making the problem less demanding. MVMED exploits multiple views in a margin consistency style. They enforced the margins from two views to be same, which means that the classification confidences from different views are considered to match each other exactly. Suppose we are given by multi-view dataset X 1 t and X 2 t, (t = 1, 2,, N) where X 1 t and X 2 t denote the ith example from view 1 and view 2, respectively and y t denotes the corresponding label. MVMED aims to seek a joint distribution p(θ 1, Θ 2 ) over the first view classifier parameter Θ 1 and the second view classifier parameter Θ 2. It uses the augmented joint distribution p(θ 1, Θ 2, γ) with the common margin γ = {γ 1,, γ N }. The optimization problem of MVMED is formulated as min p(θ1,θ 2,γ), p 0(Θ 1,Θ 2,γ) KL(p(Θ 1, Θ 2, γ) p 0 (Θ 1, Θ 2, γ)) p(θ1, Θ 2, γ) [y t L 1 (Xt 1 Θ 1 ) γ t ] dθ 1 dθ 2 dγ 0, s.t. p(θ1, Θ 2, γ) [y t L 2 (Xt 2 Θ 1 ) γ t ] dθ 1 dθ 2 dγ 0, 1 t N, where L 1 (X 1 t Θ 1 ) and L 2 (X 2 t Θ 2 ) are discriminant functions from view 1 and view 2, respectively. The expected large margin constraints are enforced on two views. The solution to the MvMED problem is based on the following theorem Theorem 1 The solution to the MVMED problem has the following general form p(θ 1, Θ 2, γ) = 1 Z(λ 1, λ 2 ) p 0(Θ 1, Θ 2, γ)e N t=1 λ1t[ytl1(x1 t Θ1) γt]+ N t=1 λ2t[ytl1(x2 t Θ2) γt], (6) where Z(λ 1, λ 2 ) is the normalization constant and λ 1 = {λ 11,..., λ 1N }, λ 2 = {λ 21,..., λ 2N } define two sets of nonnegative Lagrange multipliers, one for each classification constraint. λ 1 and λ 2 are set by finding the unique maximum of the following jointly concave objective function. That is min λ 1t,λ 2t { (5) J(λ 1, λ 2 ) = logz((λ 1, λ 2 ). (7) N (λ 1t + λ 2t + log(1 λ 1t + λ 2t )) 1 c 2 t=1 1 2 N N t=1 k=1 λ 1t λ 1k y t y k x 2 t x 2 k N N t=1 k=1 λ 1t > 0, λ 2t > 0, N t=1 λ 1ty t = 0, N t=1 λ 2ty t = 0, t {1, 2,, N}. λ 1t λ 1k y t y k x 1 t x 1 k (8)

11 Multi-view Opinion Mining with Deep Learning 11 After obtaining λ 1 and λ 2, the distribution p(θ 1, Θ 2, γ) will be specified accordingly. By marginalizing out γ, we can get the distribution p(θ 1, Θ 2 ) to predict the label of a new example (X 1, X 2 ) from view 1 and view 2 with the following rule y = sign ( 1 2 p(θ 1, Θ 2 )(L 1 (X 1 t Θ 1 ) + L 2 (X 2 t Θ 2 ))dθ 1 dθ 2 ). (9) 5 Experiment We implemented the model mentioned above, with the multi-view classification parts being MVMED, SVM-2K or CCA with kernel SVM. Experiments were conducted on 3 data sets: Large Movie Review Dataset, Stanford Sentiment Treebank and Amazon Fine Food Reviews. The split of training set, development set and test set follows the original split in Large Movie Review Dataset and Stanford Sentiment Treebank. As for Amazon Fine Food Reviews, 2.5% of the reviews are in the test set, and the remaining data are used as the training set in a 5-fold cross validation. Binary labeling criteria vary from one data set to another. Labels in Large Movie Review Dataset follow the original criteria. In Stanford Sentiment Treebank, reviews with sentiment values over 0.6 are considered as positive reviews, and those whose values are below 0.4 are negative. In Amazon Fine Food Reviews, reviews with their scores being 4 or 5 are positive, and those scored 1 or 2 are negative. All neutral reviews are ignored. Hyperparameters in each experiment are optimized through grid searching. Such hyperparameters include the unified length of documents, window sizes and filter counts in each channel in the CNN, and GRU cell counts and attention layer size in the RNN, as well as the hyperparameters in MVMED, SVM-2K and CCA with vanilla SVM. Our framework focuses on boosting performance of models given sparse data sets. From each data set, we randomly picked 1,000, 2,000, 3,000 documents respectively as the training set, while the test data set remained untouched. After training, the two DNNs (TextCNN and Attn. GRU) are individually tested, and the three multi-view models that are integrated with multi-view parts are also tested. Results are shown in Table. 1 and Fig. 6. We can conclude from the data that all multi-view methods improve the classification accuracy. They promise higher accuracies than either of the single-view methods. Specifically, CCA with SVM performs the best on Large Movie Review Dataset, and MVMED, on the other hand, performs the best on Amazon Fine Food Reviews. SVM-2K does not work well probably because it is an instance of MVMED [12], which suggests its lower generalization ability. Besides, on any one of the three data sets, accuracy increases with the growth of size of the data set. The amount of improvement is measured in the next experiment. Generally speaking, the framework leverages complementarity of heterogeneous deep features and improves overall classification accuracy on document-level opinion mining.

12 12 Ping Huang et al. Table 1 Percentage accuracies of single-view methods and multi-view methods. ACL-IMDB SST Polarity Amazon Polarity 1,000 2,000 3,000 1,000 2,000 3,000 1,000 2,000 3,000 TextCNN Attn. GRU CCA + SVM SVM-2K MVMED Accuracy(%) Training set size TextCNN Attn. GRU CCA + SVM SVM-2K MVMED Accuracy(%) Training set size TextCNN Attn. GRU CCA + SVM SVM-2K MVMED Accuracy(%) Training set size TextCNN Attn. GRU CCA + SVM SVM-2K MVMED (a) ACL-IMDB (b) SST Polarity (c) Amazon Food Polarity Fig. 6 Percentage accuracies of single-view methods and multi-view methods. Table 2 Percentage improvement from single-view methods to multi-view methods. 1,000 2,000 3,000 Avg. ACL-IMDB SST Polarity Amazon Food Polarity Avg In order to measure our model s ability to improve overall performance, we also calculated the percentage improvement of accuracies from the best single-view method to the best multi-view method in the test above. Results are shown in Table. 2. Please note that overall accuracies on Large Movie Review Dataset are the lowest among the three data sets, while those on Amazon Fine Food Reviews are the highest. From the table, we find that the lower single-view accuracies are, the more the framework can improve. So this framework helps more if applied on less reliable DNNs. 6 Conclusion Analysis on the structures of CNNs and RNNs implies that they extract different views of deep features. We explored the possibility of leveraging different features of a single document at the same time by constructing a multi-view learning framework to combine features extracted from heterogeneous DNNs. Our experiments show that, for document-level opinion mining, multi-view techniques including CCA, SVM-2K and MVMED can combine and exploit

13 Multi-view Opinion Mining with Deep Learning 13 the heterogeneous deep features and outperform single-view DNNs. Besides, our framework not only performs better as data set size grows, but also provides better improvement if applied to less reliable systems or situations in which fewer data are given. Our framework is a creative method to apply multi-view methods to any single-view problems. Further works can be done on applying multi-view methods to other conventional tasks, adapting DNNs to better fit multi-view learning, etc. 7 Acknowledgments The first two authors Ping Huang and Xijiong Xie are joint first authors. This work is sponsored by Shanghai Sailing Program. The corresponding author Shiliang Sun would also like to thank supports by NSFC Projects and , and Shanghai Knowledge Service Platform Project (No. ZF1213). The work of Xijiong Xie was supported by the NSFC of Zhejiang Province under Project LQ18F References 1. Liu B (2012) Sentiment analysis and opinion mining. Synthesis lectures on human language technologies 5(1): Sun S, Luo C, Chen J (2017) A review of natural language processing techniques for opinion mining systems. Information Fusion 36: Kalchbrenner N, Grefenstette E, Blunsom P (2014) A convolutional neural network for modelling sentences. arxiv preprint arxiv: Irsoy O, Cardie C (2014) Opinion mining with deep recurrent neural networks. In Proceedings of the conference on empirical methods in natural language processing (EMNLP), pp LeCun Y, Bengio Y, Hinton G (2015) Deep learning. nature 521(7553): Yu J, Tao D, Wang M, Rui Y (2015) Learning to rank using user clicks and visual features for image retrieval. IEEE Transactions on Cybernetics 45: Yu J, Rui Y, Tao D (2014) Click prediction for web image reranking using multimodal sparse coding. IEEE Transactions on Image Processing 23: Yu J, Yang X, Gao F, Tao D (2016) Deep multimodal distance metric learning using click constraints for image ranking. IEEE Transactions on Cybernetics 47: Tao D, Hong C, Yu J, Wan J, Wang M (2015) Multimodal deep autoencoder for human pose recovery. IEEE Transactions on Image Processing 24: Hong C,Yu J, Tao D, Wang M (2015) Image-based three-dimensional human pose recovery by multiview locality-sensitive sparse retrieval. IEEE Transactions on Industrial Electronics 62: Zhao J, Xie X, Xu X, Sun S (2017) Multi-view learning overview: Recent progress and new challenges. Information Fusion 38: Sun S, Chao G (2013) Multi-view maximum entropy discrimination. In Proceedings of the 23rd International Joint Conference on Artificial Intelligence, pp Yu j, Rui Y, Chen B (2013) Exploiting click constraints and multi-view features for image re-ranking. IEEE Transactions on Multimedia 16: Yu j, Wang M, Tao D (2012) Semi-supervised multiview distance metric learning for cartoon synthesis. IEEE Transactions on Multimedia 21: Hardoon D.R, Szedmak S.R, Shawe-Taylor J.R (2004) Canonical correlation analysis: An overview with application to learning methods. Neural Computation 12: Niu T, Zhu S, Pang L, El Saddik A (2016) Sentiment analysis on multi-view social data. In Proceedings of the International Conference on Multimedia Modeling, pp

14 14 Ping Huang et al. 17. Tang J, Hu X, Gao H, Liu H (2013) Unsupervised feature selection for multi-view data in social media. In Proceedings of the SIAM International Conference on Data Mining, SIAM, pp Wan X (2009) Co-training for cross-lingual sentiment classification. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP: Volume 1-volume 1, Association for Computational Linguistics, pp Hajmohammadi M.S, Ibrahim R, Selamat, A (2014) Bi-view semi-supervised active learning for cross-lingual sentiment classification. Information Processing & Management 50(5): Le Cun Y, Jackel L, Boser B, Denker J, Graf H, Guyon I, Henderson D, Howard R, Hubbard W (1989) Handwritten digit recognition: Applications of neural network chips and automatic learning. IEEE Communications Magazine 27: LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document recognition. In Proceedings of the IEEE 86(11): Tian DP (2013) A review on image feature extraction and representation techniques. International Journal of Multimedia and Ubiquitous Engineering 8(4): Egmont-Petersen M, de Ridder D, Handels H (2002) Image processing with neural networks a review. Pattern recognition 35(10): Schuster M, Paliwal K.K (1997) Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11): Lipton Z.C, Berkowitz J, Elkan C (2015) A critical review of recurrent neural networks for sequence learning. arxiv preprint arxiv: Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8) (1997) Cho K, Van Merriënboer B, Gulcehre C, Bahdanau D, Bougares F, Schwenk H, Bengio Y (2014) Learning phrase representations using rnn encoder-decoder for statistical machine translation. arxiv preprint arxiv: Turian J, Ratinov L, Bengio Y (2010) Word representations: a simple and general method for semi-supervised learning. In Proceedings of the 48th annual meeting of the association for computational linguistics, Association for Computational Linguistics pp Kim Y (2014) Convolutional neural networks for sentence classification. arxiv preprint arxiv: Bahdanau D, Cho K, Bengio Y (2014) Neural machine translation by jointly learning to align and translate. arxiv preprint arxiv: Farquhar J, Hardoon D, Meng H, Shawe-taylor J.S, Szedmak S (2006) Two view learning: Svm-2k, theory and practice. Advances in neural information processing systems 18: Chao G, Sun S (2016) Alternative multiview maximum entropy discrimination. IEEE transactions on neural networks and learning systems 27: Mao L, Sun S (2016) Soft margin consistency based scalable multi-view maximum entropy discrimination. In Proceedings of the 25th International Joint Conference on Artificial Intelligence, pp Maas A.L, Daly R.E, Pham, P.T, Huang D, Ng A.Y, Potts C (2011) Learning word vectors for sentiment analysis. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies-volume 1, Association for Computational Linguistics, pp

Empirical Evaluation of RNN Architectures on Sentence Classification Task

Empirical Evaluation of RNN Architectures on Sentence Classification Task Empirical Evaluation of RNN Architectures on Sentence Classification Task Lei Shen, Junlin Zhang Chanjet Information Technology lorashen@126.com, zhangjlh@chanjet.com Abstract. Recurrent Neural Networks

More information

Show, Discriminate, and Tell: A Discriminatory Image Captioning Model with Deep Neural Networks

Show, Discriminate, and Tell: A Discriminatory Image Captioning Model with Deep Neural Networks Show, Discriminate, and Tell: A Discriminatory Image Captioning Model with Deep Neural Networks Zelun Luo Department of Computer Science Stanford University zelunluo@stanford.edu Te-Lin Wu Department of

More information

End-To-End Spam Classification With Neural Networks

End-To-End Spam Classification With Neural Networks End-To-End Spam Classification With Neural Networks Christopher Lennan, Bastian Naber, Jan Reher, Leon Weber 1 Introduction A few years ago, the majority of the internet s network traffic was due to spam

More information

Sentiment Classification of Food Reviews

Sentiment Classification of Food Reviews Sentiment Classification of Food Reviews Hua Feng Department of Electrical Engineering Stanford University Stanford, CA 94305 fengh15@stanford.edu Ruixi Lin Department of Electrical Engineering Stanford

More information

Building Corpus with Emoticons for Sentiment Analysis

Building Corpus with Emoticons for Sentiment Analysis Building Corpus with Emoticons for Sentiment Analysis Changliang Li 1(&), Yongguan Wang 2, Changsong Li 3,JiQi 1, and Pengyuan Liu 2 1 Kingsoft AI Laboratory, 33, Xiaoying West Road, Beijing 100085, China

More information

Layerwise Interweaving Convolutional LSTM

Layerwise Interweaving Convolutional LSTM Layerwise Interweaving Convolutional LSTM Tiehang Duan and Sargur N. Srihari Department of Computer Science and Engineering The State University of New York at Buffalo Buffalo, NY 14260, United States

More information

Neural Network Models for Text Classification. Hongwei Wang 18/11/2016

Neural Network Models for Text Classification. Hongwei Wang 18/11/2016 Neural Network Models for Text Classification Hongwei Wang 18/11/2016 Deep Learning in NLP Feedforward Neural Network The most basic form of NN Convolutional Neural Network (CNN) Quite successful in computer

More information

LSTM and its variants for visual recognition. Xiaodan Liang Sun Yat-sen University

LSTM and its variants for visual recognition. Xiaodan Liang Sun Yat-sen University LSTM and its variants for visual recognition Xiaodan Liang xdliang328@gmail.com Sun Yat-sen University Outline Context Modelling with CNN LSTM and its Variants LSTM Architecture Variants Application in

More information

DEEP LEARNING REVIEW. Yann LeCun, Yoshua Bengio & Geoffrey Hinton Nature Presented by Divya Chitimalla

DEEP LEARNING REVIEW. Yann LeCun, Yoshua Bengio & Geoffrey Hinton Nature Presented by Divya Chitimalla DEEP LEARNING REVIEW Yann LeCun, Yoshua Bengio & Geoffrey Hinton Nature 2015 -Presented by Divya Chitimalla What is deep learning Deep learning allows computational models that are composed of multiple

More information

Machine Learning 13. week

Machine Learning 13. week Machine Learning 13. week Deep Learning Convolutional Neural Network Recurrent Neural Network 1 Why Deep Learning is so Popular? 1. Increase in the amount of data Thanks to the Internet, huge amount of

More information

Densely Connected Bidirectional LSTM with Applications to Sentence Classification

Densely Connected Bidirectional LSTM with Applications to Sentence Classification Densely Connected Bidirectional LSTM with Applications to Sentence Classification Zixiang Ding 1, Rui Xia 1(B), Jianfei Yu 2,XiangLi 1, and Jian Yang 1 1 School of Computer Science and Engineering, Nanjing

More information

Bilevel Sparse Coding

Bilevel Sparse Coding Adobe Research 345 Park Ave, San Jose, CA Mar 15, 2013 Outline 1 2 The learning model The learning algorithm 3 4 Sparse Modeling Many types of sensory data, e.g., images and audio, are in high-dimensional

More information

Learning to Match. Jun Xu, Zhengdong Lu, Tianqi Chen, Hang Li

Learning to Match. Jun Xu, Zhengdong Lu, Tianqi Chen, Hang Li Learning to Match Jun Xu, Zhengdong Lu, Tianqi Chen, Hang Li 1. Introduction The main tasks in many applications can be formalized as matching between heterogeneous objects, including search, recommendation,

More information

LSTM for Language Translation and Image Captioning. Tel Aviv University Deep Learning Seminar Oran Gafni & Noa Yedidia

LSTM for Language Translation and Image Captioning. Tel Aviv University Deep Learning Seminar Oran Gafni & Noa Yedidia 1 LSTM for Language Translation and Image Captioning Tel Aviv University Deep Learning Seminar Oran Gafni & Noa Yedidia 2 Part I LSTM for Language Translation Motivation Background (RNNs, LSTMs) Model

More information

Hierarchical Video Frame Sequence Representation with Deep Convolutional Graph Network

Hierarchical Video Frame Sequence Representation with Deep Convolutional Graph Network Hierarchical Video Frame Sequence Representation with Deep Convolutional Graph Network Feng Mao [0000 0001 6171 3168], Xiang Wu [0000 0003 2698 2156], Hui Xue, and Rong Zhang Alibaba Group, Hangzhou, China

More information

Clustering and Unsupervised Anomaly Detection with l 2 Normalized Deep Auto-Encoder Representations

Clustering and Unsupervised Anomaly Detection with l 2 Normalized Deep Auto-Encoder Representations Clustering and Unsupervised Anomaly Detection with l 2 Normalized Deep Auto-Encoder Representations Caglar Aytekin, Xingyang Ni, Francesco Cricri and Emre Aksu Nokia Technologies, Tampere, Finland Corresponding

More information

Improving Image Segmentation Quality Via Graph Theory

Improving Image Segmentation Quality Via Graph Theory International Symposium on Computers & Informatics (ISCI 05) Improving Image Segmentation Quality Via Graph Theory Xiangxiang Li, Songhao Zhu School of Automatic, Nanjing University of Post and Telecommunications,

More information

Domain-Aware Sentiment Classification with GRUs and CNNs

Domain-Aware Sentiment Classification with GRUs and CNNs Domain-Aware Sentiment Classification with GRUs and CNNs Guangyuan Piao 1(B) and John G. Breslin 2 1 Insight Centre for Data Analytics, Data Science Institute, National University of Ireland Galway, Galway,

More information

16-785: Integrated Intelligence in Robotics: Vision, Language, and Planning. Spring 2018 Lecture 14. Image to Text

16-785: Integrated Intelligence in Robotics: Vision, Language, and Planning. Spring 2018 Lecture 14. Image to Text 16-785: Integrated Intelligence in Robotics: Vision, Language, and Planning Spring 2018 Lecture 14. Image to Text Input Output Classification tasks 4/1/18 CMU 16-785: Integrated Intelligence in Robotics

More information

An efficient face recognition algorithm based on multi-kernel regularization learning

An efficient face recognition algorithm based on multi-kernel regularization learning Acta Technica 61, No. 4A/2016, 75 84 c 2017 Institute of Thermomechanics CAS, v.v.i. An efficient face recognition algorithm based on multi-kernel regularization learning Bi Rongrong 1 Abstract. A novel

More information

JOINT INTENT DETECTION AND SLOT FILLING USING CONVOLUTIONAL NEURAL NETWORKS. Puyang Xu, Ruhi Sarikaya. Microsoft Corporation

JOINT INTENT DETECTION AND SLOT FILLING USING CONVOLUTIONAL NEURAL NETWORKS. Puyang Xu, Ruhi Sarikaya. Microsoft Corporation JOINT INTENT DETECTION AND SLOT FILLING USING CONVOLUTIONAL NEURAL NETWORKS Puyang Xu, Ruhi Sarikaya Microsoft Corporation ABSTRACT We describe a joint model for intent detection and slot filling based

More information

Deep Learning Applications

Deep Learning Applications October 20, 2017 Overview Supervised Learning Feedforward neural network Convolution neural network Recurrent neural network Recursive neural network (Recursive neural tensor network) Unsupervised Learning

More information

Deep Learning. Deep Learning. Practical Application Automatically Adding Sounds To Silent Movies

Deep Learning. Deep Learning. Practical Application Automatically Adding Sounds To Silent Movies http://blog.csdn.net/zouxy09/article/details/8775360 Automatic Colorization of Black and White Images Automatically Adding Sounds To Silent Movies Traditionally this was done by hand with human effort

More information

Gate-Variants of Gated Recurrent Unit (GRU) Neural Networks

Gate-Variants of Gated Recurrent Unit (GRU) Neural Networks Gate-Variants of Gated Recurrent Unit (GRU) Neural Networks Rahul Dey and Fathi M. Salem Circuits, Systems, and Neural Networks (CSANN) LAB Department of Electrical and Computer Engineering Michigan State

More information

Facial Expression Classification with Random Filters Feature Extraction

Facial Expression Classification with Random Filters Feature Extraction Facial Expression Classification with Random Filters Feature Extraction Mengye Ren Facial Monkey mren@cs.toronto.edu Zhi Hao Luo It s Me lzh@cs.toronto.edu I. ABSTRACT In our work, we attempted to tackle

More information

A Sparse and Locally Shift Invariant Feature Extractor Applied to Document Images

A Sparse and Locally Shift Invariant Feature Extractor Applied to Document Images A Sparse and Locally Shift Invariant Feature Extractor Applied to Document Images Marc Aurelio Ranzato Yann LeCun Courant Institute of Mathematical Sciences New York University - New York, NY 10003 Abstract

More information

Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling Authors: Junyoung Chung, Caglar Gulcehre, KyungHyun Cho and Yoshua Bengio Presenter: Yu-Wei Lin Background: Recurrent Neural

More information

Image Question Answering using Convolutional Neural Network with Dynamic Parameter Prediction

Image Question Answering using Convolutional Neural Network with Dynamic Parameter Prediction Image Question Answering using Convolutional Neural Network with Dynamic Parameter Prediction by Noh, Hyeonwoo, Paul Hongsuck Seo, and Bohyung Han.[1] Presented : Badri Patro 1 1 Computer Vision Reading

More information

Recurrent Neural Nets II

Recurrent Neural Nets II Recurrent Neural Nets II Steven Spielberg Pon Kumar, Tingke (Kevin) Shen Machine Learning Reading Group, Fall 2016 9 November, 2016 Outline 1 Introduction 2 Problem Formulations with RNNs 3 LSTM for Optimization

More information

27: Hybrid Graphical Models and Neural Networks

27: Hybrid Graphical Models and Neural Networks 10-708: Probabilistic Graphical Models 10-708 Spring 2016 27: Hybrid Graphical Models and Neural Networks Lecturer: Matt Gormley Scribes: Jakob Bauer Otilia Stretcu Rohan Varma 1 Motivation We first look

More information

KBSVM: KMeans-based SVM for Business Intelligence

KBSVM: KMeans-based SVM for Business Intelligence Association for Information Systems AIS Electronic Library (AISeL) AMCIS 2004 Proceedings Americas Conference on Information Systems (AMCIS) December 2004 KBSVM: KMeans-based SVM for Business Intelligence

More information

Context Encoding LSTM CS224N Course Project

Context Encoding LSTM CS224N Course Project Context Encoding LSTM CS224N Course Project Abhinav Rastogi arastogi@stanford.edu Supervised by - Samuel R. Bowman December 7, 2015 Abstract This project uses ideas from greedy transition based parsing

More information

LSTM: An Image Classification Model Based on Fashion-MNIST Dataset

LSTM: An Image Classification Model Based on Fashion-MNIST Dataset LSTM: An Image Classification Model Based on Fashion-MNIST Dataset Kexin Zhang, Research School of Computer Science, Australian National University Kexin Zhang, U6342657@anu.edu.au Abstract. The application

More information

A FRAMEWORK OF EXTRACTING MULTI-SCALE FEATURES USING MULTIPLE CONVOLUTIONAL NEURAL NETWORKS. Kuan-Chuan Peng and Tsuhan Chen

A FRAMEWORK OF EXTRACTING MULTI-SCALE FEATURES USING MULTIPLE CONVOLUTIONAL NEURAL NETWORKS. Kuan-Chuan Peng and Tsuhan Chen A FRAMEWORK OF EXTRACTING MULTI-SCALE FEATURES USING MULTIPLE CONVOLUTIONAL NEURAL NETWORKS Kuan-Chuan Peng and Tsuhan Chen School of Electrical and Computer Engineering, Cornell University, Ithaca, NY

More information

A Sparse and Locally Shift Invariant Feature Extractor Applied to Document Images

A Sparse and Locally Shift Invariant Feature Extractor Applied to Document Images A Sparse and Locally Shift Invariant Feature Extractor Applied to Document Images Marc Aurelio Ranzato Yann LeCun Courant Institute of Mathematical Sciences New York University - New York, NY 10003 Abstract

More information

Dynamic Routing Between Capsules

Dynamic Routing Between Capsules Report Explainable Machine Learning Dynamic Routing Between Capsules Author: Michael Dorkenwald Supervisor: Dr. Ullrich Köthe 28. Juni 2018 Inhaltsverzeichnis 1 Introduction 2 2 Motivation 2 3 CapusleNet

More information

Automatic Domain Partitioning for Multi-Domain Learning

Automatic Domain Partitioning for Multi-Domain Learning Automatic Domain Partitioning for Multi-Domain Learning Di Wang diwang@cs.cmu.edu Chenyan Xiong cx@cs.cmu.edu William Yang Wang ww@cmu.edu Abstract Multi-Domain learning (MDL) assumes that the domain labels

More information

Bagging and Boosting Algorithms for Support Vector Machine Classifiers

Bagging and Boosting Algorithms for Support Vector Machine Classifiers Bagging and Boosting Algorithms for Support Vector Machine Classifiers Noritaka SHIGEI and Hiromi MIYAJIMA Dept. of Electrical and Electronics Engineering, Kagoshima University 1-21-40, Korimoto, Kagoshima

More information

Linked Document Classification by Network Representation Learning

Linked Document Classification by Network Representation Learning Linked Document Classification by Network Representation Learning Yue Zhang 1,2, Liying Zhang 1,2 and Yao Liu 1* 1 Institute of Scientific and Technical Information of China, Beijing, China 2 School of

More information

Lecture 21 : A Hybrid: Deep Learning and Graphical Models

Lecture 21 : A Hybrid: Deep Learning and Graphical Models 10-708: Probabilistic Graphical Models, Spring 2018 Lecture 21 : A Hybrid: Deep Learning and Graphical Models Lecturer: Kayhan Batmanghelich Scribes: Paul Liang, Anirudha Rayasam 1 Introduction and Motivation

More information

CSE 250B Project Assignment 4

CSE 250B Project Assignment 4 CSE 250B Project Assignment 4 Hani Altwary haltwa@cs.ucsd.edu Kuen-Han Lin kul016@ucsd.edu Toshiro Yamada toyamada@ucsd.edu Abstract The goal of this project is to implement the Semi-Supervised Recursive

More information

Support Vector Regression for Software Reliability Growth Modeling and Prediction

Support Vector Regression for Software Reliability Growth Modeling and Prediction Support Vector Regression for Software Reliability Growth Modeling and Prediction 925 Fei Xing 1 and Ping Guo 2 1 Department of Computer Science Beijing Normal University, Beijing 100875, China xsoar@163.com

More information

Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank text

Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank text Philosophische Fakultät Seminar für Sprachwissenschaft Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank text 06 July 2017, Patricia Fischer & Neele Witte Overview Sentiment

More information

Deep Learning based Authorship Identification

Deep Learning based Authorship Identification Deep Learning based Authorship Identification Chen Qian Tianchang He Rao Zhang Department of Electrical Engineering Stanford University, Stanford, CA 94305 cqian23@stanford.edu th7@stanford.edu zhangrao@stanford.edu

More information

Machine Learning. The Breadth of ML Neural Networks & Deep Learning. Marc Toussaint. Duy Nguyen-Tuong. University of Stuttgart

Machine Learning. The Breadth of ML Neural Networks & Deep Learning. Marc Toussaint. Duy Nguyen-Tuong. University of Stuttgart Machine Learning The Breadth of ML Neural Networks & Deep Learning Marc Toussaint University of Stuttgart Duy Nguyen-Tuong Bosch Center for Artificial Intelligence Summer 2017 Neural Networks Consider

More information

On the Efficiency of Recurrent Neural Network Optimization Algorithms

On the Efficiency of Recurrent Neural Network Optimization Algorithms On the Efficiency of Recurrent Neural Network Optimization Algorithms Ben Krause, Liang Lu, Iain Murray, Steve Renals University of Edinburgh Department of Informatics s17005@sms.ed.ac.uk, llu@staffmail.ed.ac.uk,

More information

Perceptron: This is convolution!

Perceptron: This is convolution! Perceptron: This is convolution! v v v Shared weights v Filter = local perceptron. Also called kernel. By pooling responses at different locations, we gain robustness to the exact spatial location of image

More information

Image Captioning with Object Detection and Localization

Image Captioning with Object Detection and Localization Image Captioning with Object Detection and Localization Zhongliang Yang, Yu-Jin Zhang, Sadaqat ur Rehman, Yongfeng Huang, Department of Electronic Engineering, Tsinghua University, Beijing 100084, China

More information

COMP 551 Applied Machine Learning Lecture 16: Deep Learning

COMP 551 Applied Machine Learning Lecture 16: Deep Learning COMP 551 Applied Machine Learning Lecture 16: Deep Learning Instructor: Ryan Lowe (ryan.lowe@cs.mcgill.ca) Slides mostly by: Class web page: www.cs.mcgill.ca/~hvanho2/comp551 Unless otherwise noted, all

More information

Visual object classification by sparse convolutional neural networks

Visual object classification by sparse convolutional neural networks Visual object classification by sparse convolutional neural networks Alexander Gepperth 1 1- Ruhr-Universität Bochum - Institute for Neural Dynamics Universitätsstraße 150, 44801 Bochum - Germany Abstract.

More information

Recurrent Neural Networks. Nand Kishore, Audrey Huang, Rohan Batra

Recurrent Neural Networks. Nand Kishore, Audrey Huang, Rohan Batra Recurrent Neural Networks Nand Kishore, Audrey Huang, Rohan Batra Roadmap Issues Motivation 1 Application 1: Sequence Level Training 2 Basic Structure 3 4 Variations 5 Application 3: Image Classification

More information

Multi-Glance Attention Models For Image Classification

Multi-Glance Attention Models For Image Classification Multi-Glance Attention Models For Image Classification Chinmay Duvedi Stanford University Stanford, CA cduvedi@stanford.edu Pararth Shah Stanford University Stanford, CA pararth@stanford.edu Abstract We

More information

Deep Learning. Deep Learning provided breakthrough results in speech recognition and image classification. Why?

Deep Learning. Deep Learning provided breakthrough results in speech recognition and image classification. Why? Data Mining Deep Learning Deep Learning provided breakthrough results in speech recognition and image classification. Why? Because Speech recognition and image classification are two basic examples of

More information

Combining SVMs with Various Feature Selection Strategies

Combining SVMs with Various Feature Selection Strategies Combining SVMs with Various Feature Selection Strategies Yi-Wei Chen and Chih-Jen Lin Department of Computer Science, National Taiwan University, Taipei 106, Taiwan Summary. This article investigates the

More information

Controllable Generative Adversarial Network

Controllable Generative Adversarial Network Controllable Generative Adversarial Network arxiv:1708.00598v2 [cs.lg] 12 Sep 2017 Minhyeok Lee 1 and Junhee Seok 1 1 School of Electrical Engineering, Korea University, 145 Anam-ro, Seongbuk-gu, Seoul,

More information

Karami, A., Zhou, B. (2015). Online Review Spam Detection by New Linguistic Features. In iconference 2015 Proceedings.

Karami, A., Zhou, B. (2015). Online Review Spam Detection by New Linguistic Features. In iconference 2015 Proceedings. Online Review Spam Detection by New Linguistic Features Amir Karam, University of Maryland Baltimore County Bin Zhou, University of Maryland Baltimore County Karami, A., Zhou, B. (2015). Online Review

More information

UTS submission to Google YouTube-8M Challenge 2017

UTS submission to Google YouTube-8M Challenge 2017 UTS submission to Google YouTube-8M Challenge 2017 Linchao Zhu Yanbin Liu Yi Yang University of Technology Sydney {zhulinchao7, csyanbin, yee.i.yang}@gmail.com Abstract In this paper, we present our solution

More information

arxiv: v1 [cs.cv] 14 Jul 2017

arxiv: v1 [cs.cv] 14 Jul 2017 Temporal Modeling Approaches for Large-scale Youtube-8M Video Understanding Fu Li, Chuang Gan, Xiao Liu, Yunlong Bian, Xiang Long, Yandong Li, Zhichao Li, Jie Zhou, Shilei Wen Baidu IDL & Tsinghua University

More information

Deep Multi-Modal Learning with Cascade Consensus

Deep Multi-Modal Learning with Cascade Consensus Deep Multi-Modal Learning with Cascade Consensus Yang Yang, Yi-Feng Wu, De-Chuan Zhan, Yuan Jiang National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China {yangy,

More information

Deep Learning. Vladimir Golkov Technical University of Munich Computer Vision Group

Deep Learning. Vladimir Golkov Technical University of Munich Computer Vision Group Deep Learning Vladimir Golkov Technical University of Munich Computer Vision Group 1D Input, 1D Output target input 2 2D Input, 1D Output: Data Distribution Complexity Imagine many dimensions (data occupies

More information

Aggregating Frame-level Features for Large-Scale Video Classification

Aggregating Frame-level Features for Large-Scale Video Classification Aggregating Frame-level Features for Large-Scale Video Classification Shaoxiang Chen 1, Xi Wang 1, Yongyi Tang 2, Xinpeng Chen 3, Zuxuan Wu 1, Yu-Gang Jiang 1 1 Fudan University 2 Sun Yat-Sen University

More information

10-701/15-781, Fall 2006, Final

10-701/15-781, Fall 2006, Final -7/-78, Fall 6, Final Dec, :pm-8:pm There are 9 questions in this exam ( pages including this cover sheet). If you need more room to work out your answer to a question, use the back of the page and clearly

More information

GraphNet: Recommendation system based on language and network structure

GraphNet: Recommendation system based on language and network structure GraphNet: Recommendation system based on language and network structure Rex Ying Stanford University rexying@stanford.edu Yuanfang Li Stanford University yli03@stanford.edu Xin Li Stanford University xinli16@stanford.edu

More information

Hello Edge: Keyword Spotting on Microcontrollers

Hello Edge: Keyword Spotting on Microcontrollers Hello Edge: Keyword Spotting on Microcontrollers Yundong Zhang, Naveen Suda, Liangzhen Lai and Vikas Chandra ARM Research, Stanford University arxiv.org, 2017 Presented by Mohammad Mofrad University of

More information

The Un-normalized Graph p-laplacian based Semi-supervised Learning Method and Speech Recognition Problem

The Un-normalized Graph p-laplacian based Semi-supervised Learning Method and Speech Recognition Problem Int. J. Advance Soft Compu. Appl, Vol. 9, No. 1, March 2017 ISSN 2074-8523 The Un-normalized Graph p-laplacian based Semi-supervised Learning Method and Speech Recognition Problem Loc Tran 1 and Linh Tran

More information

Image-Sentence Multimodal Embedding with Instructive Objectives

Image-Sentence Multimodal Embedding with Instructive Objectives Image-Sentence Multimodal Embedding with Instructive Objectives Jianhao Wang Shunyu Yao IIIS, Tsinghua University {jh-wang15, yao-sy15}@mails.tsinghua.edu.cn Abstract To encode images and sentences into

More information

Deep Learning on Graphs

Deep Learning on Graphs Deep Learning on Graphs with Graph Convolutional Networks Hidden layer Hidden layer Input Output ReLU ReLU, 22 March 2017 joint work with Max Welling (University of Amsterdam) BDL Workshop @ NIPS 2016

More information

XES Tensorflow Process Prediction using the Tensorflow Deep-Learning Framework

XES Tensorflow Process Prediction using the Tensorflow Deep-Learning Framework XES Tensorflow Process Prediction using the Tensorflow Deep-Learning Framework Demo Paper Joerg Evermann 1, Jana-Rebecca Rehse 2,3, and Peter Fettke 2,3 1 Memorial University of Newfoundland 2 German Research

More information

arxiv: v1 [cs.cv] 4 Feb 2018

arxiv: v1 [cs.cv] 4 Feb 2018 End2You The Imperial Toolkit for Multimodal Profiling by End-to-End Learning arxiv:1802.01115v1 [cs.cv] 4 Feb 2018 Panagiotis Tzirakis Stefanos Zafeiriou Björn W. Schuller Department of Computing Imperial

More information

Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting

Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting Yaguang Li Joint work with Rose Yu, Cyrus Shahabi, Yan Liu Page 1 Introduction Traffic congesting is wasteful of time,

More information

Character Recognition Using Convolutional Neural Networks

Character Recognition Using Convolutional Neural Networks Character Recognition Using Convolutional Neural Networks David Bouchain Seminar Statistical Learning Theory University of Ulm, Germany Institute for Neural Information Processing Winter 2006/2007 Abstract

More information

Semi-Supervised Clustering with Partial Background Information

Semi-Supervised Clustering with Partial Background Information Semi-Supervised Clustering with Partial Background Information Jing Gao Pang-Ning Tan Haibin Cheng Abstract Incorporating background knowledge into unsupervised clustering algorithms has been the subject

More information

Semi-supervised learning and active learning

Semi-supervised learning and active learning Semi-supervised learning and active learning Le Song Machine Learning II: Advanced Topics CSE 8803ML, Spring 2012 Combining classifiers Ensemble learning: a machine learning paradigm where multiple learners

More information

Lecture 7: Neural network acoustic models in speech recognition

Lecture 7: Neural network acoustic models in speech recognition CS 224S / LINGUIST 285 Spoken Language Processing Andrew Maas Stanford University Spring 2017 Lecture 7: Neural network acoustic models in speech recognition Outline Hybrid acoustic modeling overview Basic

More information

Learning Meanings for Sentences with Recursive Autoencoders

Learning Meanings for Sentences with Recursive Autoencoders Learning Meanings for Sentences with Recursive Autoencoders Tsung-Yi Lin and Chen-Yu Lee Department of Electrical and Computer Engineering University of California, San Diego {tsl008, chl260}@ucsd.edu

More information

CSC 578 Neural Networks and Deep Learning

CSC 578 Neural Networks and Deep Learning CSC 578 Neural Networks and Deep Learning Fall 2018/19 7. Recurrent Neural Networks (Some figures adapted from NNDL book) 1 Recurrent Neural Networks 1. Recurrent Neural Networks (RNNs) 2. RNN Training

More information

Face Recognition Based on LDA and Improved Pairwise-Constrained Multiple Metric Learning Method

Face Recognition Based on LDA and Improved Pairwise-Constrained Multiple Metric Learning Method Journal of Information Hiding and Multimedia Signal Processing c 2016 ISSN 2073-4212 Ubiquitous International Volume 7, Number 5, September 2016 Face Recognition ased on LDA and Improved Pairwise-Constrained

More information

PTE : Predictive Text Embedding through Large-scale Heterogeneous Text Networks

PTE : Predictive Text Embedding through Large-scale Heterogeneous Text Networks PTE : Predictive Text Embedding through Large-scale Heterogeneous Text Networks Pramod Srinivasan CS591txt - Text Mining Seminar University of Illinois, Urbana-Champaign April 8, 2016 Pramod Srinivasan

More information

Bidirectional Recurrent Convolutional Networks for Video Super-Resolution

Bidirectional Recurrent Convolutional Networks for Video Super-Resolution Bidirectional Recurrent Convolutional Networks for Video Super-Resolution Qi Zhang & Yan Huang Center for Research on Intelligent Perception and Computing (CRIPAC) National Laboratory of Pattern Recognition

More information

Rotation Invariance Neural Network

Rotation Invariance Neural Network Rotation Invariance Neural Network Shiyuan Li Abstract Rotation invariance and translate invariance have great values in image recognition. In this paper, we bring a new architecture in convolutional neural

More information

Learning to Rank with Attentive Media Attributes

Learning to Rank with Attentive Media Attributes Learning to Rank with Attentive Media Attributes Baldo Faieta Yang (Allie) Yang Adobe Adobe San Francisco, CA 94103 San Francisco, CA. 94103 bfaieta@adobe.com yangyan@adobe.com Abstract In the context

More information

One-class Problems and Outlier Detection. 陶卿 中国科学院自动化研究所

One-class Problems and Outlier Detection. 陶卿 中国科学院自动化研究所 One-class Problems and Outlier Detection 陶卿 Qing.tao@mail.ia.ac.cn 中国科学院自动化研究所 Application-driven Various kinds of detection problems: unexpected conditions in engineering; abnormalities in medical data,

More information

VisoLink: A User-Centric Social Relationship Mining

VisoLink: A User-Centric Social Relationship Mining VisoLink: A User-Centric Social Relationship Mining Lisa Fan and Botang Li Department of Computer Science, University of Regina Regina, Saskatchewan S4S 0A2 Canada {fan, li269}@cs.uregina.ca Abstract.

More information

Recurrent Neural Networks

Recurrent Neural Networks Recurrent Neural Networks Javier Béjar Deep Learning 2018/2019 Fall Master in Artificial Intelligence (FIB-UPC) Introduction Sequential data Many problems are described by sequences Time series Video/audio

More information

Deep MIML Network. Abstract

Deep MIML Network. Abstract Deep MIML Network Ji Feng and Zhi-Hua Zhou National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China Collaborative Innovation Center of Novel Software Technology

More information

learning stage (Stage 1), CNNH learns approximate hash codes for training images by optimizing the following loss function:

learning stage (Stage 1), CNNH learns approximate hash codes for training images by optimizing the following loss function: 1 Query-adaptive Image Retrieval by Deep Weighted Hashing Jian Zhang and Yuxin Peng arxiv:1612.2541v2 [cs.cv] 9 May 217 Abstract Hashing methods have attracted much attention for large scale image retrieval.

More information

Show, Discriminate, and Tell: A Discriminatory Image Captioning Model with Deep Neural Networks

Show, Discriminate, and Tell: A Discriminatory Image Captioning Model with Deep Neural Networks Show, Discriminate, and Tell: A Discriminatory Image Captioning Model with Deep Neural Networks Boya Peng Department of Computer Science Stanford University boya@stanford.edu Zelun Luo Department of Computer

More information

Proceedings of the International MultiConference of Engineers and Computer Scientists 2018 Vol I IMECS 2018, March 14-16, 2018, Hong Kong

Proceedings of the International MultiConference of Engineers and Computer Scientists 2018 Vol I IMECS 2018, March 14-16, 2018, Hong Kong , March 14-16, 2018, Hong Kong , March 14-16, 2018, Hong Kong , March 14-16, 2018, Hong Kong , March 14-16, 2018, Hong Kong TABLE I CLASSIFICATION ACCURACY OF DIFFERENT PRE-TRAINED MODELS ON THE TEST DATA

More information

TRANSDUCTIVE TRANSFER LEARNING BASED ON KL-DIVERGENCE. Received January 2013; revised May 2013

TRANSDUCTIVE TRANSFER LEARNING BASED ON KL-DIVERGENCE. Received January 2013; revised May 2013 International Journal of Innovative Computing, Information and Control ICIC International c 2014 ISSN 1349-4198 Volume 10, Number 1, February 2014 pp. 303 313 TRANSDUCTIVE TRANSFER LEARNING BASED ON KL-DIVERGENCE

More information

Opinion Mining by Transformation-Based Domain Adaptation

Opinion Mining by Transformation-Based Domain Adaptation Opinion Mining by Transformation-Based Domain Adaptation Róbert Ormándi, István Hegedűs, and Richárd Farkas University of Szeged, Hungary {ormandi,ihegedus,rfarkas}@inf.u-szeged.hu Abstract. Here we propose

More information

Code Mania Artificial Intelligence: a. Module - 1: Introduction to Artificial intelligence and Python:

Code Mania Artificial Intelligence: a. Module - 1: Introduction to Artificial intelligence and Python: Code Mania 2019 Artificial Intelligence: a. Module - 1: Introduction to Artificial intelligence and Python: 1. Introduction to Artificial Intelligence 2. Introduction to python programming and Environment

More information

Cost-alleviative Learning for Deep Convolutional Neural Network-based Facial Part Labeling

Cost-alleviative Learning for Deep Convolutional Neural Network-based Facial Part Labeling [DOI: 10.2197/ipsjtcva.7.99] Express Paper Cost-alleviative Learning for Deep Convolutional Neural Network-based Facial Part Labeling Takayoshi Yamashita 1,a) Takaya Nakamura 1 Hiroshi Fukui 1,b) Yuji

More information

Keyword Extraction by KNN considering Similarity among Features

Keyword Extraction by KNN considering Similarity among Features 64 Int'l Conf. on Advances in Big Data Analytics ABDA'15 Keyword Extraction by KNN considering Similarity among Features Taeho Jo Department of Computer and Information Engineering, Inha University, Incheon,

More information

String Vector based KNN for Text Categorization

String Vector based KNN for Text Categorization 458 String Vector based KNN for Text Categorization Taeho Jo Department of Computer and Information Communication Engineering Hongik University Sejong, South Korea tjo018@hongik.ac.kr Abstract This research

More information

Texture classification using convolutional neural networks

Texture classification using convolutional neural networks University of Wollongong Research Online Faculty of Informatics - Papers (Archive) Faculty of Engineering and Information Sciences 2006 Texture classification using convolutional neural networks Fok Hing

More information

Deep Learning in Visual Recognition. Thanks Da Zhang for the slides

Deep Learning in Visual Recognition. Thanks Da Zhang for the slides Deep Learning in Visual Recognition Thanks Da Zhang for the slides Deep Learning is Everywhere 2 Roadmap Introduction Convolutional Neural Network Application Image Classification Object Detection Object

More information

Link Prediction for Social Network

Link Prediction for Social Network Link Prediction for Social Network Ning Lin Computer Science and Engineering University of California, San Diego Email: nil016@eng.ucsd.edu Abstract Friendship recommendation has become an important issue

More information

Neural Network and Deep Learning. Donglin Zeng, Department of Biostatistics, University of North Carolina

Neural Network and Deep Learning. Donglin Zeng, Department of Biostatistics, University of North Carolina Neural Network and Deep Learning Early history of deep learning Deep learning dates back to 1940s: known as cybernetics in the 1940s-60s, connectionism in the 1980s-90s, and under the current name starting

More information

Neural Network Joint Language Model: An Investigation and An Extension With Global Source Context

Neural Network Joint Language Model: An Investigation and An Extension With Global Source Context Neural Network Joint Language Model: An Investigation and An Extension With Global Source Context Ruizhongtai (Charles) Qi Department of Electrical Engineering, Stanford University rqi@stanford.edu Abstract

More information

CS231N Section. Video Understanding 6/1/2018

CS231N Section. Video Understanding 6/1/2018 CS231N Section Video Understanding 6/1/2018 Outline Background / Motivation / History Video Datasets Models Pre-deep learning CNN + RNN 3D convolution Two-stream What we ve seen in class so far... Image

More information