AAAI 2021 论文列表

Training deep neural networks using a large batch size has shown promising results and benefits many real-world applications. Warmup is one of nontrivial techniques to stabilize the convergence of large batch training. However, warmup is an empirical method and it is still unknown whether there is a better algorithm with theoretical underpinnings. In this paper, we propose a novel Complete Layer-wise Adaptive Rate Scaling (CLARS) algorithm for large-batch training. We prove the convergence of our algorithm by introducing a new fine-grained analysis of gradient-based methods. Furthermore, the new analysis also helps to understand two other empirical tricks, layer-wise adaptive rate scaling and linear learning rate scaling. We conduct extensive experiments and demonstrate that the proposed algorithm outperforms gradual warmup technique by a large margin and defeats the convergence of the state-of-the-art large-batch optimizer in training advanced deep neural networks (ResNet, DenseNet, MobileNet) on ImageNet dataset.

Landmarks (LMs) are state features that need to be made true or tasks that need to be contained in every solution of a planning problem. They are a valuable source of information in planning and can be exploited in various ways. LMs have been used both in classical and hierarchical planning, but while there is much work in classical planning, the techniques in hierarchical planning are less evolved. We introduce a novel LM generation method for Hierarchical Task Network (HTN) planning and show that it is sound and incomplete. We show that every complete approach is as hard as the co-class of the underlying HTN problem, i.e. coNP-hard for our setting (while our approach is in P). On a widely used benchmark set, our approach finds more than twice the number of landmarks than the approach from the literature. Though our focus is on LM generation, we show that the newly discovered landmarks bear information beneficial for solvers.

Representing the true label as one-hot vector is the common practice in training text classification models. However, the one-hot representation may not adequately reflect the relation between the instance and labels, as labels are often not completely independent and instances may relate to multiple labels in practice. The inadequate one-hot representations tend to train the model to be over-confident, which may result in arbitrary prediction and model overfitting, especially for confused datasets (datasets with very similar labels) or noisy datasets (datasets with labeling errors). While training models with label smoothing can ease this problem in some degree, it still fails to capture the realistic relation among labels. In this paper, we propose a novel Label Confusion Model (LCM) as an enhancement component to current popular text classification models. LCM can learn label confusion to capture semantic overlap among labels by calculating the similarity between instance and labels during training and generate a better label distribution to replace the original one-hot label vector, thus improving the final classification performance. Extensive experiments on five text classification benchmark datasets reveal the effectiveness of LCM for several widely used deep learning classification models. Further experiments also verify that LCM is especially helpful for confused or noisy datasets and superior to the label smoothing method.

Deep learning based subspace clustering methods have attracted increasing attention in recent years, where a basic theme is to non-linearly map data into a latent space, and then uncover subspace structures based upon the data self-expressiveness property. However, almost all existing deep subspace clustering methods only rely on target domain data, and always resort to shallow neural networks for modeling data, leaving huge room to design more effective representation learning mechanisms tailored for subspace clustering. In this paper, we propose a novel subspace clustering framework through learning precise sample representations. In contrast to previous approaches, the proposed method aims to leverage external data through constructing lots of relevant tasks to guide the training of the encoder, motivated by the idea of meta-learning. Considering limited layer structures of current deep subspace clustering models, we intend to distill knowledge from a deeper network trained on the external data, and transfer it into the shallower model. To reach the above two goals, we propose a new loss function to realize them in a joint framework. Moreover, we propose to construct a new pretext task for self-supervised training of the model, such that the representation ability of the model can be further improved. Extensive experiments are performed on four publicly available datasets, and experimental results clearly demonstrate the efficacy of our method, compared to state-of-the-art methods.

Hyperspectral anomaly detection (HAD) is a challenging task because it explores the intrinsic structure of complex high-dimensional signals without any samples at training time. Deep neural networks (DNNs) can dig out the underlying distribution of hyperspectral data but are limited by the labeling of large-scale hyperspectral datasets, especially the low spatial resolution of hyperspectral data, which makes labeling more difficult. To tackle this problem while ensuring the detection performance, we present an unsupervised low-rank embedded network (LREN) in this paper. LREN is a joint learning network in which the latent representation is specifically designed for HAD, rather than merely as a feature input for the detector. And it searches the lowest rank representation based on a representative and discriminative dictionary in the deep latent space to estimate the residual efficiently. Considering the physically mixing properties in hyperspectral imaging, we develop a trainable density estimation module based on Gaussian mixture model (GMM) in the deep latent space to construct a dictionary that can better characterize the complex hyperspectral images (HSIs). The closed-form solution of the proposed low-rank learner surpasses existing approaches on four real hyperspectral datasets with different anomalies. We argue that this unified framework paves a novel way to combine feature extraction and anomaly estimation-based methods for HAD, which intends to learn the underlying representation tailored for HAD without the prerequisite of manually labeled data. Code available at https://github.com/xdjiangkai/LREN.

The pre-training models such as BERT have achieved great results in various natural language processing problems. However, a large number of parameters need significant amounts of memory and the consumption of inference time, which makes it difficult to deploy them on edge devices. In this work, we propose a knowledge distillation method LRC-BERT based on contrastive learning to fit the output of the intermediate layer from the angular distance aspect, which is not considered by the existing distillation methods. Furthermore, we introduce a gradient perturbation-based training architecture in the training phase to increase the robustness of LRC-BERT, which is the first attempt in knowledge distillation. Additionally, in order to better capture the distribution characteristics of the intermediate layer, we design a two-stage training method for the total distillation loss. Finally, by verifying 8 datasets on the General Language Understanding Evaluation (GLUE) benchmark, the performance of the proposed LRC-BERT exceeds the existing state-of-the-art methods, which proves the effectiveness of our method.

Natural language explanations (NLEs) are a special form of data annotation in which annotators identify rationales (most significant text tokens) when assigning labels to data instances, and write out explanations for the labels in natural language based on the rationales. NLEs have been shown to capture human reasoning better, but not as beneficial for natural language inference (NLI). In this paper, we analyze two primary flaws in the way NLEs are currently used to train explanation generators for language inference tasks. We find that the explanation generators do not take into account the variability inherent in human explanation of labels, and that the current explanation generation models generate spurious explanations. To overcome these limitations, we propose a novel framework, LIREx, that incorporates both a rationale-enabled explanation generator and an instance selector to select only relevant, plausible NLEs to augment NLI models. When evaluated on the standardized SNLI data set, LIREx achieved an accuracy of 91.87%, an improvement of 0.32 over the baseline and matching the best-reported performance on the data set. It also achieves significantly better performance than previous studies when transferred to the out-of-domain MultiNLI data set. Qualitative analysis shows that LIREx generates flexible, faithful, and relevant NLEs that allow the model to be more robust to spurious explanations. The code is available at https://github.com/zhaoxy92/LIREx.

Chinese short text matching is a fundamental task in natural language processing. Existing approaches usually take Chinese characters or words as input tokens. They have two limitations: 1) Some Chinese words are polysemous, and semantic information is not fully utilized. 2) Some models suffer potential issues caused by word segmentation. Here we introduce HowNet as an external knowledge base and propose a Linguistic knowledge Enhanced graph Transformer (LET) to deal with word ambiguity. Additionally, we adopt the word lattice graph as input to maintain multi-granularity information. Our model is also complementary to pre-trained language models. Experimental results on two Chinese datasets show that our models outperform various typical text matching approaches. Ablation study also indicates that both semantic information and multi-granularity information are important for text matching modeling.

We present LCollision, a learning-based method that synthesizes collision-free 3D human poses. At the crux of our approach is a novel deep architecture that simultaneously decodes new human poses from the latent space and predicts colliding body parts. These two components of our architecture are used as the objective function and surrogate hard constraints in a constrained optimization for collision-free human pose generation. A novel aspect of our approach is the use of a bilevel autoencoder that decomposes whole-body collisions into groups of collisions between localized body parts. By solving the constrained optimizations, we show that a significant amount of collision artifacts can be resolved. Furthermore, in a large test set of 2.5 × 10 6 randomized poses from SCAPE, our architecture achieves a collision-prediction accuracy of 94.1% with 80× speedup over exact collision detection algorithms. To the best of our knowledge, LCollision is the first approach that accelerates collision detection and resolves penetrations using a neural network.

Understanding the meaning of a text is a fundamental challenge of natural language understanding (NLU) research. An ideal NLU system should process a language in a way that is not exclusive to a single task or a dataset. Keeping this in mind, we have introduced a novel knowledge driven semantic representation approach for English text. By leveraging the VerbNet lexicon, we are able to map syntax tree of the text to its commonsense meaning represented using basic knowledge primitives. The general purpose knowledge represented from our approach can be used to build any reasoning based NLU system that can also provide justification. We applied this approach to construct two NLU applications that we present here: SQuARE (Semantic-based Question Answering and Reasoning Engine) and StaCACK (Stateful Conversational Agent using Commonsense Knowledge). Both these systems work by ``truly understanding'' the natural language text they process and both provide natural language explanations for their responses while maintaining high accuracy.

Recent developments in pre-trained neural language modeling have led to leaps in accuracy on common-sense question-answering benchmarks. However, there is increasing concern that models overfit to specific tasks, without learning to utilize external knowledge or perform general semantic reasoning. In contrast, zero-shot evaluations have shown promise as a more robust measure of a model’s general reasoning abilities. In this paper, we propose a novel neuro-symbolic framework for zero-shot question answering across commonsense tasks. Guided by a set of hypotheses, the framework studies how to transform various pre-existing knowledge resources into a form that is most effective for pre-training models. We vary the set of language models, training regimes, knowledge sources, and data generation strategies, and measure their impact across tasks. Extending on prior work, we devise and compare four constrained distractor-sampling strategies. We provide empirical results across five commonsense question-answering tasks with data generated from five external knowledge resources. We show that, while an individual knowledge graph is better suited for specific tasks, a global knowledge graph brings consistent gains across different tasks. In addition, both preserving the structure of the task as well as generating fair and informative questions help language models learn more effectively.

Named Entity Recognition (NER) is a fundamental and important research topic for many downstream NLP tasks, aiming at detecting and classifying named entities (NEs) mentioned in unstructured text into pre-defined categories. Learning from labeled data only is far from enough when it comes to domain-specific or temporally-evolving entities (medical terminologies or restaurant names). Luckily, open-source Knowledge Bases (KBs) (Wikidata and Freebase) contain NEs that are manually labeled with predefined types in different domains, which is potentially beneficial to identify entity boundaries and recognize entity types more accurately. However, the type system of a domain-specific NER task is typically independent of that of current KBs and thus exhibits heterogeneity issue inevitably, which makes matching between the original NER and KB types (Person in NER potentially matches President in KBs) less likely, or introduces unintended noises without considering domain-specific knowledge (Band in NER should be mapped to Out_of_Entity_Types in the restaurant-related task). To better incorporate and denoise the abundant knowledge in KBs, we propose a new KB-aware NER framework (KaNa), which utilizes type-heterogeneous knowledge to improve NER. Specifically, for an entity mention along with a set of candidate entities that are linked from KBs, KaNa first uses a type projection mechanism that maps the mention type and entity types into a shared space to homogenize the heterogeneous entity types. Then, based on projected types, a noise detector filters out certain less-confident candidate entities in an unsupervised manner. Finally, the filtered mention-entity pairs are injected into a NER model as a graph to predict answers. The experimental results demonstrate KaNa's state-of-the-art performance on five public benchmark datasets from different domains.

While widely used in industry, recurrent neural networks (RNNs) are known to have deficiencies in dealing with long sequences (e.g. slow inference, vanishing gradients etc.). Recent research has attempted to accelerate RNN models by developing mechanisms to skip irrelevant words in input. Due to the lack of labelled data, it remains as a challenge to decide which words to skip, especially for low-resource classification tasks. In this paper, we propose Knowledge-AwareLeap-LSTM (KALL), a novel architecture which integrates prior human knowledge (created either manually or automatically) like in-domain keywords, terminologies or lexicons into Leap-LSTM to partially supervise the skipping process. More specifically, we propose a knowledge-oriented cost function for KALL; furthermore, we propose two strategies to integrate the knowledge: (1) the Factored KALL approach involves a keyword indicator as a soft constraint for the skip-ping process, and (2) the Gated KALL enforces the inclusion of keywords while maintaining a differentiable network in training. Experiments on different public datasets show that our approaches are1.1x~2.6x faster than LSTM with better accuracy and 23.6x faster than XLNet in a resource-limited CPU-only environment.

Social recommendation task aims to predict users' preferences over items with the incorporation of social connections among users, so as to alleviate the sparse issue of collaborative filtering. While many recent efforts show the effectiveness of neural network-based social recommender systems, several important challenges have not been well addressed yet: (i) The majority of models only consider users’ social connections, while ignoring the inter-dependent knowledge across items; (ii) Most of existing solutions are designed for singular type of user-item interactions, making them infeasible to capture the interaction heterogeneity; (iii) The dynamic nature of user-item interactions has been less explored in many social-aware recommendation techniques. To tackle the above challenges, this work proposes a Knowledge-aware Coupled Graph Neural Network (KCGN) that jointly injects the inter-dependent knowledge across items and users into the recommendation framework. KCGN enables the high-order user- and item-wise relation encoding by exploiting the mutual information for global graph structure awareness. Additionally, we further augment KCGN with the capability of capturing dynamic multi-typed user-item interactive patterns. Experimental studies on real-world datasets show the effectiveness of our method against many strong baselines in a variety of settings. Source codes are available at: https://github.com/xhcdream/KCGN.

We present a framework called Acquired Deep Impressions (ADI) which continuously learns knowledge of objects as ``impressions'' for compositional scene understanding. In this framework, the model first acquires knowledge from scene images containing a single object in a supervised manner, and then continues to learn from novel multi-object scene images which may contain objects that have not been seen before without any further supervision, under the guidance of the learned knowledge as humans do. By memorizing impressions of objects into parameters of neural networks and applying the generative replay strategy, the learned knowledge can be reused to generate images with pseudo-annotations and in turn assist the learning of novel scenes. The proposed ADI framework focuses on the acquisition and utilization of knowledge, and is complementary to existing deep generative models proposed for compositional scene representation. We adapt a base model to make it fall within the ADI framework and conduct experiments on two types of datasets. Empirical results suggest that the proposed framework is able to effectively utilize the acquired impressions and improve the scene decomposition performance.

Personalized recommender systems are increasingly important as more content and services become available and users struggle to identify what might interest them. Thanks to the ability for providing rich information, knowledge graphs (KGs) are being incorporated to enhance the recommendation performance and interpretability. To effectively make use of the knowledge graph, we propose a recommendation model in the hyperbolic space, which facilitates the learning of the hierarchical structure of knowledge graphs. Furthermore, a hyperbolic attention network is employed to determine the relative importances of neighboring entities of a certain item. In addition, we propose an adaptive and fine-grained regularization mechanism to adaptively regularize items and their neighboring representations. Via a comparison using three real-world datasets with state-of-the-art methods, we show that the proposed model outperforms the best existing models by 2-16% in terms of NDCG@K on Top-K recommendation.

Accurate user and item embedding learning is crucial for modern recommender systems. However, most existing recommendation techniques have thus far focused on modeling users' preferences over singular type of user-item interactions. Many practical recommendation scenarios involve multi-typed user interactive behaviors (e.g., page view, add-to-favorite and purchase), which presents unique challenges that cannot be handled by current recommendation solutions. In particular: i) complex inter-dependencies across different types of user behaviors; ii) the incorporation of knowledge-aware item relations into the multi-behavior recommendation framework; iii) dynamic characteristics of multi-typed user-item interactions. To tackle these challenges, this work proposes a Knowledge-Enhanced Hierarchical Graph Transformer Network (KHGT), to investigate multi-typed interactive patterns between users and items in recommender systems. Specifically, KHGT is build upon a graph-structured neural architecture to i) capture type-specific behavior semantics; ii) explicitly discriminate which types of user-item interactions are more important in assisting the forecasting task on the target behavior. Additionally, we further integrate the multi-modal graph attention layer with temporal encoding strategy, to empower the learned embeddings be reflective of both dedicated multiplex user-item and item-item collaborative relations, as well as the underlying interaction dynamics. Extensive experiments conducted on three real-world datasets show that KHGT consistently outperforms many state-of-the-art recommendation methods across various evaluation settings. Our implementation is available in https://github.com/akaxlh/KHGT.

In this paper, we propose a novel configurable framework to automatically generate distractive choices for open-domain cloze-style multiple-choice questions. The framework incorporates a general-purpose knowledge base to effectively create a small distractor candidate set, and a feature-rich learning-to-rank model to select distractors that are both plausible and reliable. Experimental results on a new dataset across four domains show that our framework yields distractors outperforming previous methods both by automatic and human evaluation. The dataset can also be used as a benchmark for distractor generation research in the future.

Description logics (DLs) are knowledge representation languages that are used in the field of artificial intelligence (AI). A common technique is to query DL knowledge bases, e.g., by Boolean Datalog queries, and ask for entailment. But real world knowledge-bases are often obtained by combining data from various sources. This, inherently, might result in certain inconsistencies (with respect to a given query) and requires to estimate a degree of inconsistency before using a knowledge-base. In this paper, we provide a complexity analysis of fixed-domain non-entailment (NE) on Datalog programs for well-established families of knowledge bases (KBs). We exhibit a detailed complexity map for the decision cases, counting and projected counting, which may serve as a quantitative measure for inconsistency of a KB with respect to a query. Our results show that NE is natural for the second, third, and fourth level of the polynomial (counting) hierarchy depending on the type of the studied query (stratified, normal, disjunctive) and one level higher for the projected versions. Further, we show fixed-parameter tractability by bounding the treewidth, provide a constructive algorithm, and show its theoretical limitation in terms of conditional lower bounds.