Scenario description:The AI summit IJCAI 2019 was successfully concluded on august 16. In the 7-day technical pageant, the participants learned about the application scenarios of AI technology in various fields in the workshop, listened to the keynote speeches of senior AI experts, and also had the opportunity to learn about the historical story of AI development and the latest progress and trends in the round table. In addition, the papers included in the meeting are undoubtedly the most concerned content. We hereby organize several selected papers in different fields to share with you.
Key words:IJCAI 2019 paper
The top-level conference on artificial intelligence IJCAI 2019 was held from August 10 to August 16 in Macao, China and ended successfully.
At the opening ceremony on August 13, the organizers of the conference combed the collection of papers for this conference. Thomas Eiter, president of the conference, announced the following information: IJCAI received a total of 4,752 papers for submission this year, and the final number of papers collected reached a record high of 850, with an acceptance rate of 17.9%.
Then Sarit Kraus, chairman of the procedural committee of the conference, gave a detailed description of the papers. Compared with 3,470 papers collected last year, this year’s growth rate was 37%, and 327 of the 850 papers collected came from China, accounting for 38%.
On the theme of the paper, machine learning is still the hottest field, with 438 articles included, more than half of them. In addition, the fields with the largest number of papers are computer vision, machine learning application and natural language processing in turn.
Of the papers submitted, 2516 are in the field of machine learning.
This year, 73 field chairmen, 740 members of the high-level procedural committee and 2,696 members of the procedural committee participated in the paper review. What are the outstanding papers reviewed by them?
One-in-a-Hundred Award-winning Thesis
IJCAI 2019 selects a Distinguished Paper from 850 papers:
Abstract: The author studies a classification problem based on comparison. This kind of problem is generally like this: given a set, only triplet information can be obtained. this triplet information is the comparison of three targets. for example, the distance from x_i to x_j is smaller than the distance from x_i to x_k, how can x_i be classified? In this paper, researchers proposed an algorithm called “TripletBoost” to learn classifiers from such triplet data. The main idea of the paper is that the distance information brought by triplets can be input into a weak classifier, which can be serialized and gradually upgraded to a strong classifier.
This method has two advantages: firstly, this method can be applied in various matrix spaces; in addition, this method can solve triplet information that can only be obtained passively or noise in many fields. The researchers theoretically verified the feasibility of this method in the paper and put forward the lower limit of the number of triplets that need to be obtained. Through experiments, they said that this method is better than the existing methods and can resist noise better.
The Best Paper for IJCAI-JAIR is:
Address of the paper:https://www.jair.org/index.ph …
Note: This award is given to papers published in JAIR in the past five years.
This paper indicates that the most typical NP-complete problem (NP-complete), boolean SATisfiability (sat) and its generalized version of PSPACE-complete quantitative boolean satisfiability (QAT) are the core of declarative programming paradigm, which can efficiently solve various computationally complex problems in real world examples.
The success in this field is achieved through breakthroughs in the practical implementation of SAT and QSAT decision programs, namely SAT and QSAT solvers. In this paper, researchers have developed and analyzed the clause elimination process for preprocessing and post-processing. Among them, the clause elimination process forms a set of (P)CNF formal simplification techniques, so clauses with specific redundancy characteristics are removed in polynomial time while keeping the formula satisfying.
In addition to these award-winning papers, IJCAI, as one of the hottest top conferences in the field of artificial intelligence, ranks among the top conferences in terms of both the number of papers submitted and the number of papers received over the years. There are many outstanding papers among them.
Therefore, supranerve will select one or two selected papers from the three hottest fields (Hot 3) of the current IJCAI conference for a brief introduction, so as to see the spot and leopard and the whole picture of IJCAI.
Hot 1 in the hottest field: machine learning
Selected Papers on Machine Learning 1
Address of the paper:
The Open Video Question Answering is to automatically generate text answers from referenced video content according to given questions.
At present, the existing methods often use multi-modal cyclic codec network, but it lacks long-term dependency modeling, which makes it unable to be effectively applied to long-term video Q&A.
In order to solve this problem, the author proposes a fast hierarchical convolutional self-attention codec network (HCSA). Using a self-attention encoder with layered convolution, long video content is effectively modeled.
HCSA establishes a hierarchical structure of video sequences and captures long-term dependencies with problem awareness from the video context. In addition, a multi-scale attention decoder is designed, which integrates multi-layer representation to generate the answer and avoids the information loss of the top coding layer.
Experimental results show that the method performs well on multiple data sets.
Selected Papers on Machine Learning 2
Address of the paper:
Absrtact: The application of machine learning is often limited by the number of effective labeled data, and semi-supervised learning can effectively solve this problem.
This paper proposes a simple and effective semi-supervised learning algorithm, interpolation consistency training (ICT).
ICT makes the interpolation prediction of unlabeled points consistent with the interpolation prediction of these points. In classification problem, ICT moves the decision boundary to the low density area of data distribution. It requires little extra computation and no training to generate models. Even without extensive super-parameter tuning, it achieves the most advanced performance when applied to standard neural network architectures on CIFAR-10 and SVHN benchmark data sets.
Hot 2 in the hottest field: computer vision
Selected Papers on Computer Vision 1
Address of the paper:
Features from multiple scales can greatly help semantic edge detection tasks. However, common semantic edge detection methods apply a fixed weight fusion strategy, in which images with different semantics are forced to share the same weight, resulting in common fusion weights for all images and locations, regardless of their different semantics or local contexts.
This work proposes a novel dynamic feature fusion strategy, which adaptively assigns different fusion weights to different input images and positions. This is achieved by a suggested weight learner to infer appropriate fusion weights for specific inputs with multi-level features of each location of the feature map.
In this way, the heterogeneity of contributions made by different positions of the feature map and the input image can be better considered, thus contributing to more accurate and clearer edge prediction.
Selected Papers on Computer Vision 2
Address of the paper:
Absrtact: Monocular depth estimation is an important task in scene understanding. Objects and their underlying structures in complex scenes are very important for accurate restoration of depth maps with good visual effects. The global structure reflects the scene layout and the local structure reflects the shape details. The depth estimation method based on CNN developed in recent years significantly improves the performance of depth estimation. However, multi-scale structures in complex scenes are rarely considered.
In this paper, a structure-aware residual pyramid network (SARPN) using multi-scale structure for accurate depth prediction is proposed, and a residual pyramid decoder (RPD) is also proposed, which represents the global scene structure to represent the layout in the upper layer and the local structure to represent the shape details in the lower layer. At each layer, a residual refinement module (RRM) of prediction residual mapping is proposed to gradually add finer structures to the coarse structures predicted at the upper layer. In order to make full use of the features of multi-scale images, an adaptive dense feature fusion (ADFF) module is proposed, which adaptively fuses the effective features of various scales and is used for reasoning on the structures of various scales. The experimental results on NYU-Depth v2 data set show that the proposed method achieves the most advanced performance in both qualitative and quantitative evaluation, with precision of 0.749, recall rate of 0.554 and F1 score of 0.630.
Hot 3: Natural Language Processing (NLP) in Hot Fields
NLP Selected Papers 1
Address of the paper:
Abstract: Recurrent Neural Networks (RNNs) are widely used in the field of Natural Language Processing (NLP), including text classification, question answering and machine translation. In general, RNNs can only review from beginning to end, and its ability to process long text is poor. However, in the task of text classification, a large number of words in long documents are irrelevant and can be skipped. In view of this situation, the author proposes enhanced LSTM: LEAP-LSTM.
Leap-LSTM can dynamically jump between words when reading text. In each step Leap-LSTM uses several feature encoders to extract information from the preceding text, the following text and the current word, and then decides whether to skip the current word. On the five benchmark data sets of AGNews, DBPedia, Yelp F. Yelp P. and Yahoo, the prediction effect of Leap-LSTM is higher than that of standard LSTM, and Leap-LSTM has higher reading speed.
NLP Selected Papers 2
Address of the paper:
Abstract: This paper studies the problem of entity alignment based on knowledge map embedding. Previous work focused on the relationship structure of entities, and some further combined other types of features, such as attributes, to refine them.
However, a large number of entity features are still not treated equally, which impairs the accuracy and robustness of entity alignment based on embedding.
This paper proposes a new framework, which unifies multiple views of entities to learn the embedding of entity alignment. Specifically, this paper uses several combination strategies to embed entities based on views of entity names, relationships, and attributes.
In addition, this paper designs some cross-knowledge map inference methods to enhance the alignment between the two knowledge maps. Experiments on real data sets show that the performance of this framework is significantly better than the most advanced entity alignment method based on embedding. The selected view, cross-knowledge map reasoning and combination strategy all contribute to the performance improvement.