AAAI 2021 论文列表

Controlling the style of natural language by disentangling the latent space is an important step towards interpretable machine learning. After the latent space is disentangled, the style of a sentence can be transformed by tuning the style representation without affecting other features of the sentence. Previous works usually use adversarial training to guarantee that disentangled vectors do not affect each other. However, adversarial methods are difficult to train. Especially when there are multiple features (e.g., sentiment, or tense, which we call style types in this paper), each feature requires a separate discriminator for extracting a disentangled style vector corresponding to that feature. In this paper, we propose a unified distribution-controlling method, which provides each specific style value (the value of style types, e.g., positive sentiment, or past tense) with a unique representation. This method contributes a solid theoretical basis to avoid adversarial training in multi-type disentanglement. We also propose multiple loss functions to achieve a style-content disentanglement as well as a disentanglement among multiple style types. In addition, we observe that if two different style types always have some specific style values that occur together in the dataset, they will affect each other when transferring the style values. We call this phenomenon training bias , and we propose a loss function to alleviate such training bias while disentangling multiple types. We conduct experiments on two datasets (Yelp service reviews and Amazon product reviews) to evaluate the style-disentangling effect and the unsupervised style-transfer performance on two style types: sentiment and tense. The experimental results show the effectiveness of our model.

One crucial objective of multi-task learning is to align distributions across tasks so that the information between them can be transferred and shared. However, existing approaches only focused on matching the marginal feature distribution while ignoring the semantic information, which may hinder the learning performance. To address this issue, we propose to leverage the label information in multi-task learning by exploring the semantic conditional relations among tasks. We first theoretically analyze the generalization bound of multi-task learning based on the notion of Jensen-Shannon divergence, which provides new insights into the value of label information in multi-task learning. Our analysis also leads to a concrete algorithm that jointly matches the semantic distribution and controls label distribution divergence. To confirm the effectiveness of the proposed method, we first compare the algorithm with several baselines on some benchmarks and then test the algorithms under label space shift conditions. Empirical results demonstrate that the proposed method could outperform most baselines and achieve state-of-the-art performance, particularly showing the benefits under the label shift conditions.

The key challenge of co-saliency detection is to extract discriminative features to distinguish the common salient foregrounds from backgrounds in a group of relevant images. In this paper, we propose a new co-saliency detection framework which includes two strategies to improve the discriminative ability of the features. Specifically, on one hand, we segment each image to semantic superpixel clusters as well as generate different scales/sizes of images for each input image by the VGG-16 model. Different scales capture different patterns of the images. As a result, multi-scale images can capture various patterns among all images by many kinds of perspectives. Second, we propose a new method of Graph Convolutional Network (GCN) to fine-tune the multi-scale features, aiming at capturing the common information among the features from all scales and the private or complementary information for the feature of each scale. Moreover, the proposed GCN method jointly conducts multi-scale feature fine-tune, graph learning, and feature learning in a unified framework. We evaluated our method on three benchmark data sets, compared to state-of-the-art co-saliency detection methods. Experimental results showed that our method outperformed all comparison methods in terms of different evaluation metrics.

As an important research issue in affective computing community, multi-modal emotion recognition has become a hot topic in the last few years. However, almost all existing studies perform multiple binary classification for each emotion with focus on complete time series data. In this paper, we focus on multi-modal emotion recognition in a multi-label scenario. In this scenario, we consider not only the label-to-label dependency, but also the feature-to-label and modality-to-label dependencies. Particularly, we propose a heterogeneous hierarchical message passing network to effectively model above dependencies. Furthermore, we propose a new multi-modal multi-label emotion dataset based on partial time-series content to show predominant generalization of our model. Detailed evaluation demonstrates the effectiveness of our approach.

Multi-modal named entity recognition (MNER) aims to discover named entities in free text and classify them into pre-defined types with images. However, dominant MNER models do not fully exploit fine-grained semantic correspondences between semantic units of different modalities, which have the potential to refine multi-modal representation learning. To deal with this issue, we propose a unified multi-modal graph fusion (UMGF) approach for MNER. Specifically, we first represent the input sentence and image using a unified multi-modal graph, which captures various semantic relationships between multi-modal semantic units (words and visual objects). Then, we stack multiple graph-based multi-modal fusion layers that iteratively perform semantic interactions to learn node representations. Finally, we achieve an attention-based multi-modal representation for each word and perform entity labeling with a CRF decoder. Experimentation on the two benchmark datasets demonstrates the superiority of our MNER model.

We propose a novel distance-based regularization method for deep metric learning called Multi-level Distance Regularization (MDR). MDR explicitly disturbs a learning procedure by regularizing pairwise distances between embedding vectors into multiple levels that represents a degree of similarity between a pair. In the training stage, the model is trained with both MDR and an existing loss function of deep metric learning, simultaneously; the two losses interfere with the objective of each other, and it makes the learning process difficult. Moreover, MDR prevents some examples from being ignored or overly influenced in the learning process. These allow the parameters of the embedding network to be settle on a local optima with better generalization. Without bells and whistles, MDR with simple Triplet loss achieves the-state-of-the-art performance in various benchmark datasets: CUB-200-2011, Cars-196, Stanford Online Products, and In-Shop Clothes Retrieval. We extensively perform ablation studies on its behaviors to show the effectiveness of MDR. By easily adopting our MDR, the previous approaches can be improved in performance and generalization ability.

Multi-view representation learning attempts to learn a representation from multiple views and most existing methods are unsupervised. However, representation learned only from unlabeled data may not be discriminative enough for further applications (e.g., clustering and classification). For this reason, semi-supervised methods which could use unlabeled data along with the labeled data for multi-view representation learning need to be developed. Manifold information plays an important role in semi-supervised learning, but it has not been considered for multi-view representation learning. In this paper, we introduce the manifold smoothness into multi-view representation learning and propose MvDGAT which learns the representation and the intrinsic manifold simultaneously with graph attention network. Experiments conducted on real-world datasets reveal that our MvDGAT can achieve better performance than state-of-the-art methods.

In real-world applications, clustering or classification can usually be improved by fusing information from different views. Therefore, unsupervised representation learning on multi-view data becomes a compelling topic in machine learning. In this paper, we propose a novel and flexible unsupervised multi-view representation learning model termed Collaborative Multi-View Information Bottleneck Networks (CMIB-Nets), which comprehensively explores the common latent structure and the view-specific intrinsic information, and discards the superfluous information in the data significantly improving the generalization capability of the model. Specifically, our proposed model relies on the information bottleneck principle to integrate the shared representation among different views and the view-specific representation of each view, prompting the multi-view complete representation and flexibly balancing the complementarity and consistency among multiple views. We conduct extensive experiments (including clustering analysis, robustness experiment, and ablation study) on real-world datasets, which empirically show promising generalization ability and robustness compared to state-of-the-arts.

Neural dialogue models suffer from low-quality responses when interacted in practice, demonstrating difficulty in generalization beyond training data. Recently, knowledge distillation has been used to successfully regularize the student by transferring knowledge from the teacher. However, the teacher and the student are trained on the same dataset and tend to learn similar feature representations, whereas the most general knowledge should be found through differences. The finding of general knowledge is further hindered by the unidirectional distillation, as the student should obey the teacher and may discard some knowledge that is truly general but refuted by the teacher. To this end, we propose a novel training framework, where the learning of general knowledge is more in line with the idea of reaching consensus, i.e., finding common knowledge that is beneficial to different yet all datasets through diversified learning partners. Concretely, the training task is divided into a group of subtasks with the same number of students. Each student assigned to one subtask not only is optimized on the allocated subtask but also imitates multi-view feature representation aggregated from other students (i.e., student peers), which induces students to capture common knowledge among different subtasks and alleviates the over-fitting of students on the allocated subtasks. To further enhance generalization, we extend the unidirectional distillation to the bidirectional distillation that encourages the student and its student peers to co-evolve by exchanging complementary knowledge with each other. Empirical results and analysis demonstrate that our training framework effectively improves the model generalization without sacrificing training efficiency.

We consider the models of deep multi-task learning with recurrent architectures that exploit regularities across tasks to improve the performance of multiple sequence processing tasks jointly. Most existing architectures are painstakingly customized to learn task relationships for different problems, which is not flexible enough to model the dynamic task relationships and lacks generalization abilities to novel test-time scenarios. We propose multi-task recurrent modular networks (MT-RMN) that can be incorporated in any multi-task recurrent models to address the above drawbacks. MT-RMN consists of a shared encoder and multiple task-specific decoders, and recurrently operates over time. For better flexibility, it modularizes the encoder into multiple layers of sub-networks and dynamically controls the connection between these sub-networks and the decoders at different time steps, which provides the recurrent networks with varying degrees of parameter sharing for tasks with dynamic relatedness. For the generalization ability, MT-RMN aims to discover a set of generalizable sub-networks in the encoder that are assembled in different ways for different tasks. The policy networks augmented with the differentiable routers are utilized to make the binary connection decisions between the sub-networks. The experimental results on three multi-task sequence processing datasets consistently demonstrate the effectiveness of MT-RMN.

While generative adversarial networks (GANs) based neural text-to-speech (TTS) systems have shown significant improvement in neural speech synthesis, there is no TTS system to learn to synthesize speech from text sequences with only adversarial feedback. Because adversarial feedback alone is not sufficient to train the generator, current models still require the reconstruction loss compared with the ground-truth and the generated mel-spectrogram directly. In this paper, we present Multi-SpectroGAN (MSG), which can train the multi-speaker model with only the adversarial feedback by conditioning a self-supervised hidden representation of the generator to a conditional discriminator. This leads to better guidance for generator training. Moreover, we also propose adversarial style combination (ASC) for better generalization in the unseen speaking style and transcript, which can learn latent representations of the combined style embedding from multiple mel-spectrograms. Trained with ASC and feature matching, the MSG synthesizes a high-diversity mel-spectrogram by controlling and mixing the individual speaking styles (e.g., duration, pitch, and energy). The result shows that the MSG synthesizes a high-fidelity mel-spectrogram, which has almost the same naturalness MOS score as the ground-truth mel-spectrogram.

Graph convolutional networks have been widely used for skeleton-based action recognition due to their excellent modeling ability of non-Euclidean data. As the graph convolution is a local operation, it can only utilize the short-range joint dependencies and short-term trajectory but fails to directly model the distant joints relations and long-range temporal information that are vital to distinguishing various actions. To solve this problem, we present a multi-scale spatial graph convolution (MS-GC) module and a multi-scale temporal graph convolution (MT-GC) module to enrich the receptive field of the model in spatial and temporal dimensions. Concretely, the MS-GC and MT-GC modules decompose the corresponding local graph convolution into a set of sub-graph convolution, forming a hierarchical residual architecture. Without introducing additional parameters, the features will be processed with a series of sub-graph convolutions, and each node could complete multiple spatial and temporal aggregations with its neighborhoods. The final equivalent receptive field is accordingly enlarged, which is capable of capturing both short- and long-range dependencies in spatial and temporal domains. By coupling these two modules as a basic block, we further propose a multi-scale spatial temporal graph convolutional network (MST-GCN), which stacks multiple blocks to learn effective motion representations for action recognition. The proposed MST-GCN achieves remarkable performance on three challenging benchmark datasets, NTU RGB+D, NTU-120 RGB+D and Kinetics-Skeleton, for skeleton-based action recognition.

Network games provide a natural machinery to compactly represent strategic interactions among agents whose payoffs exhibit sparsity in their dependence on the actions of others. Besides encoding interaction sparsity, however, real networks often exhibit a multi-scale structure, in which agents can be grouped into communities, those communities further grouped, and so on, and where interactions among such groups may also exhibit sparsity. We present a general model of multi-scale network games that encodes such multi-level structure. We then develop several algorithmic approaches that leverage this multi-scale structure, and derive sufficient conditions for convergence of these to a Nash equilibrium. Our numerical experiments demonstrate that the proposed approaches enable orders of magnitude improvements in scalability when computing Nash equilibria in such games. For example, we can solve previously intractable instances involving up to 1 million agents in under 15 minutes.

Most widely-used convolutional neural networks (CNNs) end up with a global average pooling layer and a fully-connected layer. In this pipeline, a certain class is represented by one template vector preserved in the feature banks of fully-connected layer. Yet, a class may have multiple properties useful for recognition while the above formulation only captures one of them. Therefore, it is desired to represent a class by multiple proxies. However, directly adding multiple linear layers turns out to be a trivial solution as no improvement can be observed. To tackle this problem, we adopt optimal transport theory to calculate a non-uniform matching flow between the elements in the feature map of a sample and the proxies of a class in a closed way. By doing so, the models are enabled to achieve partial matching as both the feature maps and the proxy set can now focus on a subset of elements from the counterpart. Such formulation also enables us to embed the samples into the Wasserstein metric space, which has many advantages over the original Euclidean space. This formulation can be achieved by a lightweight iterative algorithm, which can be easily embedded into the automatic differentiation framework. Empirical studies are performed on two widely-used classification datasets, CIFAR, and ILSVRC2012, and the substantial improvements on these two benchmarks demonstrate the effectiveness of our method.

We study a social choice setting of manipulation in elections and extend the usual model in two major ways: first, instead of considering a single manipulating agent, in our setting there are several, possibly competing ones; second, instead of evaluating an election after the first manipulative action, we allow several back-and-forth rounds to take place. We show that in certain situations, such as in elections with only a few candidates, optimal strategies for each of the manipulating agents can be computed efficiently. Our algorithmic results rely on formulating the problem of finding an optimal strategy as sentences of Presburger arithmetic that are short and only involve small coefficients, which we show is fixed-parameter tractable -- indeed, one of our contributions is a general result regarding fixed-parameter tractability of Presburger arithmetic that might be useful in other settings. Following our general theorem, we design quite general algorithms; in particular, we describe how to design efficient algorithms for various settings, including settings in which we model diffusion of opinions in a social network, complex budgeting schemes available to the manipulating agents, and various realistic restrictions on adversary actions.

Submodular maximization has attracted much attention due to its wide application and attractive property. Previous works mainly considered one single objective function, while there can be multiple ones in practice. As the objectives are usually conflicting, there exists a set of Pareto optimal solutions, attaining different optimal trade-offs among multiple objectives. In this paper, we consider the problem of minimizing the regret ratio in multi-objective submodular maximization, which is to find at most k solutions to approximate the whole Pareto set as well as possible. We propose a new algorithm RRMS by sampling representative weight vectors and solving the corresponding weighted sums of objective functions using some given alpha-approximation algorithm for single-objective submodular maximization. We prove that the regret ratio of the output of RRMS is upper bounded by 1-alpha+O(sqrt{d-1}cdot(frac{d}{k-d})^{frac{1}{d-1}}), where d is the number of objectives. This is the first theoretical guarantee for the situation with more than two objectives. When d=2, it reaches the (1-alpha+O(1/k))-guarantee of the only existing algorithm Polytope. Empirical results on the applications of multi-objective weighted maximum coverage and Max-Cut show the superior performance of RRMS over Polytope.

We introduce multi-goal multi agent path finding (MG-MAPF) which generalizes the standard discrete multi-agent path finding (MAPF) problem. While the task in MAPF is to navigate agents in an undirected graph from their starting vertices to one individual goal vertex per agent, MG-MAPF assigns each agent multiple goal vertices and the task is to visit each of them at least once. Solving MG-MAPF not only requires finding collision free paths for individual agents but also determining the order of visiting agent's goal vertices so that common objectives like the sum-of-costs are optimized. We suggest two novel algorithms using different paradigms to address MG-MAPF: a heuristic search-based algorithm called Hamiltonian-CBS (HCBS) and a compilation-based algorithm built using the satisfiability modulo theories (SMT), called SMT-Hamiltonian-CBS (SMT-HCBS).

Rehearsal, seeking to remind the model by storing old knowledge in lifelong learning, is one of the most effective ways to mitigate catastrophic forgetting, i.e., biased forgetting of previous knowledge when moving to new tasks. However, the old tasks of the most previous rehearsal-based methods suffer from the unpredictable domain shift when training the new task. This is because these methods always ignore two significant factors. First, the Data Imbalance between the new task and old tasks that makes the domain of old tasks prone to shift. Second, the Task Isolation among all tasks will make the domain shift toward unpredictable directions; To address the unpredictable domain shift, in this paper, we propose Multi-Domain Multi-Task (MDMT) rehearsal to train the old tasks and new task parallelly and equally to break the isolation among tasks. Specifically, a two-level angular margin loss is proposed to encourage the intra-class/task compactness and inter-class/task discrepancy, which keeps the model from domain chaos. In addition, to further address domain shift of the old tasks, we propose an optional episodic distillation loss on the memory to anchor the knowledge for each old task. Experiments on benchmark datasets validate the proposed approach can effectively mitigate the unpredictable domain shift.

Personality detection aims to identify the personality traits implied in social media posts. The core of this task is to put together information in multiple scattered posts to depict an overall personality profile for each user. Existing approaches either encode each post individually or assemble posts arbitrarily into a new document that can be encoded sequentially or hierarchically. While the first approach ignores the connection between posts, the second tends to introduce unnecessary post-order bias into posts. In this paper, we propose a multi-document Transformer, namely Transformer-MD, to tackle the above issues. When encoding each post, Transformer-MD allows access to information in the other posts of the user through Transformer-XL’s memory tokens which share the same position embedding.Besides, personality is usually defined along different traits and each trait may need to attend to different post information, which has rarely been touched by existing research. To address this concern, we propose a dimension attention mechanism on top of Transformer-MD to obtain trait-specific representations for multi-trait personality detection. We evaluate the proposed model on the Kaggle and Pandora MBTI datasets and the experimental results show that it compares favorably with baseline methods.