IJCAI 2022 Papers

Recently, neural techniques have been used to generate source code automatically. While promising for declarative languages, these approaches achieve much poorer performance on datasets for imperative languages. Since a declarative language is typically embedded in an imperative language (i.e., the turducken-style programming) in real-world software development, the promising results on declarative languages can hardly lead to significant reduction of manual software development efforts. In this paper, we define a new code generation task: given a natural language comment, this task aims to generate a program in a base imperative language with an embedded declarative language. To our knowledge, this is the first turducken-style code generation task. For this task, we present Lyra: a dataset in Python with embedded SQL. This dataset contains 2,000 carefully annotated database manipulation programs from real usage projects. Each program is paired with both a Chinese comment and an English comment. In our experiment, we adopted Transformer, BERT-style, and GPT-style models as baselines. In the best setting, GPT-style model can achieve 24% and 25.5% AST exact matching accuracy using Chinese and English comments, respectively. Therefore, we believe that Lyra provides a new challenge for code generation. Yet, overcoming this challenge may significantly boost the applicability of code generation techniques for real-world software development.

Named entity recognition (NER) is a fundamental information extraction task that seeks to identify entity mentions of certain types in text. Despite numerous advances, the existing NER methods rely on extensive supervision for model training, which struggle in a low-resource scenario with limited training data. In this paper, we propose a new data augmentation method for low-resource NER, by eliciting knowledge from BERT with prompting strategies. Particularly, we devise a label-conditioned word replacement strategy that can produce more label-consistent examples by capturing the underlying word-label dependencies, and a prompting with question answering method to generate new training data from unlabeled texts. The experimental results have widely confirmed the effectiveness of our approach. Particularly, in a low-resource scenario with only 150 training sentences, our approach outperforms previous methods without data augmentation by over 40% in F1 and prior best data augmentation methods by over 2.0% in F1. Furthermore, our approach also fits with a zero-shot scenario, yielding promising results without using any human-labeled data for the task.

Many real-world ubiquitous applications, such as parking recommendations and air pollution monitoring, benefit significantly from accurate long-term spatio-temporal forecasting (LSTF). LSTF makes use of long-term dependency structure between the spatial and temporal domains, as well as the contextual information. Recent studies have revealed the potential of multi-graph neural networks (MGNNs) to improve prediction performance. However, existing MGNN methods do not work well when applied to LSTF due to several issues: the low level of generality, insufficient use of contextual information, and the imbalanced graph fusion approach. To address these issues, we construct new graph models to represent the contextual information of each node and exploit the long-term spatio-temporal data dependency structure. To aggregate the information across multiple graphs, we propose a new dynamic multi-graph fusion module to characterize the correlations of nodes within a graph and the nodes across graphs via the spatial attention and graph attention mechanisms. Furthermore, we introduce a trainable weight tensor to indicate the importance of each node in different graphs. Extensive experiments on two large-scale datasets demonstrate that our proposed approaches significantly improve the performance of existing graph neural network models in LSTF prediction tasks.

Compared with image few-shot learning, most of the existing few-shot video classification methods perform worse on feature matching, because they fail to sufficiently exploit the temporal information and relation. Specifically, frames are usually evenly sampled, which may miss important frames. On the other hand, the heuristic model simply encodes the equally treated frames in sequence, which results in the lack of both long-term and short-term temporal modeling and interaction. To alleviate these limitations, we take advantage of the compressed domain knowledge and propose a long-short term Cross-Transformer (LSTC) for few-shot video classification. For short terms, the motion vector (MV) contains temporal cues and reflects the importance of each frame. For long terms, a video can be natively divided into a sequence of GOPs (Group Of Picture). Using this compressed domain knowledge helps to obtain a more accurate spatial-temporal feature space. Consequently, we design the long-short term selection module, short-term module, and long-term module to comprise the LSTC. Long-short term selection is performed to select informative compressed domain data. Long/short-term modules are utilized to sufficiently exploit the temporal information so that the query and support can be well-matched by cross-attention. Experimental results show the superiority of our method on various datasets.

When a person solves the multi-choice problem, she considers not only what is the answer but also what is not the answer. Knowing what choice is not the answer and utilizing the relationships between choices, she can improve the prediction accuracy. Inspired by this human reasoning process, we propose a new training strategy to fully utilize inter-class relationships, namely LogitMix. Our strategy is combined with recent data augmentation techniques, e.g., Mixup, Manifold Mixup, CutMix, and PuzzleMix. Then, we suggest using a mixed logit, i.e., a mixture of two logits, as an auxiliary training objective. Since the logit can preserve both positive and negative inter-class relationships, it can impose a network to learn the probability of wrong answers correctly. Our extensive experimental results on the image- and language-based tasks demonstrate that LogitMix achieves state-of-the-art performance among recent data augmentation techniques regarding calibration error and prediction accuracy.

Recent efforts within the AI community have yielded impressive results towards “soft theorem proving” over natural language sentences using language models. We propose a novel, generative adversarial framework for probing and improving these models’ reasoning capabilities. Adversarial attacks in this domain suffer from the logical inconsistency problem, whereby perturbations to the input may alter the label. Our Logically consistent AdVersarial Attacker, LAVA, addresses this by combining a structured generative process with a symbolic solver, guaranteeing logical consistency. Our framework successfully generates adversarial attacks and identifies global weaknesses common across multiple target models. Our analyses reveal naive heuristics and vulnerabilities in these models’ reasoning capabilities, exposing an incomplete grasp of logical deduction under logic programs. Finally, in addition to effective probing of these models, we show that training on the generated samples improves the target model’s performance.

The shipping industry is an important component of the global trade and economy. In order to ensure law compliance and safety, it needs to be monitored. In this paper, we present a novel ship type classification model that combines vessel transmitted data from the Automatic Identification System, with vessel imagery. The main components of our approach are the Faster R-CNN Deep Neural Network and a Neuro-Fuzzy system with IF-THEN rules. We evaluate our model using real world data and showcase the advantages of this combination while also compare it with other methods. Results show that our model can increase prediction scores by up to 15.4% when compared with the next best model we considered, while also maintaining a level of explainability as opposed to common black box approaches.

Recent deep clustering algorithms take advantage of self-supervised learning and self-training techniques to map the original data into a latent space, where the data embedding and clustering assignment can be jointly optimized. However, as many recent datasets are enormous and noisy, getting a clear boundary between different clusters is challenging with existing methods that mainly focus on contracting similar samples together and overlooking samples near boundary of clusters in the latent space. In this regard, we propose an end-to-end deep clustering algorithm, i.e., Locally Normalized Soft Contrastive Clustering (LNSCC). It takes advantage of similarities among each sample's local neighborhood and globally disconnected samples to leverage positiveness and negativeness of sample pairs in a contrastive way to separate different clusters. Experimental results on various datasets illustrate that our proposed approach achieves outstanding clustering performance over most of the state-of-the-art clustering methods for both image and non-image data even without convolution.

With advances in reinforcement learning (RL), agents are now being developed in high-stakes application domains such as healthcare and transportation. Explaining the behavior of these agents is challenging, as they act in large state spaces, and their decision-making can be affected by delayed rewards. In this paper, we explore a combination of explanations that attempt to convey the global behavior of the agent and local explanations which provide information regarding the agent's decision-making in a particular state. Specifically, we augment strategy summaries that demonstrate the agent's actions in a range of states with saliency maps highlighting the information it attends to. Our user study shows that intelligently choosing what states to include in the summary (global information) results in an improved analysis of the agents. We find mixed results with respect to augmenting summaries with saliency maps (local information).

The resilience of a voting system has been a central topic in computational social choice. Many voting rules, like plurality, are shown to be vulnerable as the attacker can target specific voters to manipulate the result. What if a local differential privacy (LDP) mechanism is adopted such that the true preference of a voter is never revealed in pre-election polls? In this case, the attacker can only infer stochastic information about a voter's true preference, and this may cause the manipulation of the electoral result significantly harder. The goal of this paper is to provide a quantitative study on the effect of adopting LDP mechanisms on a voting system. We introduce the metric PoLDP (power of LDP) that quantitatively measures the difference between the attacker's manipulation cost under LDP mechanisms and that without LDP mechanisms. The larger PoLDP is, the more robustness LDP mechanisms can add to a voting system. We give a full characterization of PoLDP for the voting system with plurality rule and provide general guidance towards the application of LDP mechanisms.

This paper studies Linear Temporal Logic over Finite Traces (LTLf) where proposition letters are replaced with first-order formulas interpreted over arbitrary theories, in the spirit of Satisfiability Modulo Theories. The resulting logic, called LTLf Modulo Theories (LTLfMT), is semi-decidable. Nevertheless, its high expressiveness comes useful in a number of use cases, such as model-checking of data-aware processes and data-aware planning. Despite the general undecidability of these problems, being able to solve satisfiable instances is a compromise worth studying. After motivating and describing such use cases, we provide a sound and complete semi-decision procedure for LTLfMT based on the SMT encoding of a one-pass tree-shaped tableau system. The algorithm is implemented in the BLACK satisfiability checking tool, and an experimental evaluation shows the feasibility of the approach on novel benchmarks.

In a sequential decision-making problem, having a structural dependency amongst the reward distributions associated with the arms makes it challenging to identify a subset of alternatives that guarantees the optimal collective outcome. Thus, besides individual actions' reward, learning the causal relations is essential to improve the decision-making strategy. To solve the two-fold learning problem described above, we develop the 'combinatorial semi-bandit framework with causally related rewards', where we model the causal relations by a directed graph in a stationary structural equation model. The nodal observation in the graph signal comprises the corresponding base arm's instantaneous reward and an additional term resulting from the causal influences of other base arms' rewards. The objective is to maximize the long-term average payoff, which is a linear function of the base arms' rewards and depends strongly on the network topology. To achieve this objective, we propose a policy that determines the causal relations by learning the network's topology and simultaneously exploits this knowledge to optimize the decision-making process. We establish a sublinear regret bound for the proposed algorithm. Numerical experiments using synthetic and real-world datasets demonstrate the superior performance of our proposed method compared to several benchmarks.

The topic of forgetting has been extensively studied in the field of knowledge representation and reasoning for many major formalisms. Quite recently it has been introduced to abstract argumentation. However, many already known as well as essential aspects about forgetting like strong persistence or strong invariance have been left unconsidered. We show that forgetting in abstract argumentation cannot be reduced to forgetting in logic programming. In addition, we deal with the more general problem of forgetting whole sets of arguments and show that iterative application of existing operators for single arguments does not necessarily yield a desirable result as it may not produce an informationally economic argumentation framework. As a consequence we provide a systematic and exhaustive study of forgetting desiderata and associated operations adapted to the intrinsics of abstract argumentation. We show the limits and shed light on the possibilities.

Single-image super-resolution (SISR) has achieved significant breakthroughs with the development of deep learning. However, these methods are difficult to be applied in real-world scenarios since they are inevitably accompanied by the problems of computational and memory costs caused by the complex operations. To solve this issue, we propose a Lightweight Bimodal Network (LBNet) for SISR. Specifically, an effective Symmetric CNN is designed for local feature extraction and coarse image reconstruction. Meanwhile, we propose a Recursive Transformer to fully learn the long-term dependence of images thus the global information can be fully used to further refine texture details. Studies show that the hybrid of CNN and Transformer can build a more efficient model. Extensive experiments have proved that our LBNet achieves more prominent performance than other state-of-the-art methods with a relatively low computational cost and memory consumption. The code is available at https://github.com/IVIPLab/LBNet.

Agent-based crawlers are commonly used in network maintenance and information gathering. In order not to disturb the main functionality of the system, whether acting at nodes or being in transit, they need to operate online, perform a single operation fast and use small memory. They should also be preferably deterministic, as crawling agents have limited capabilities of generating a large number of truly random bits. We consider a system in which an agent receives an update, typically an insertion or deletion, of some information upon visiting a node. On request, the agent needs to output hot information, i.e., with the net occurrence above certain frequency threshold. A desired time and memory complexity of such agent should be poly-logarithmic in the number of visited nodes and inversely proportional to the frequency threshold. Ours is the first such agent with rigorous analysis and a complementary almost-matching lower bound.

Universal application of AI has increased the need to explain why an AI model makes a specific decision in a human-understandable form. Among many related works, the class activation map (CAM)-based methods have been successful recently, creating input attribution based on the weighted sum of activation maps in convolutional neural networks. However, existing methods use channel-wise importance weights with specific architectural assumptions, relying on arbitrarily chosen attribution threshold values in their quality assessment: we think these can degrade the quality of attribution. In this paper, we propose Libra-CAM, a new CAM-style attribution method based on the best linear approximation of the layer (as a function) between the penultimate activation and the target-class score output. From the approximation, we derive the base formula of Libra-CAM, which is applied with multiple reference activations from a pre-built library. We construct Libra-CAM by averaging these base attribution maps, taking a threshold calibration procedure to optimize its attribution quality. Our experiments show that Libra-CAM can be computed in a reasonable time and is superior to the existing attribution methods in quantitative and qualitative attribution quality evaluations.

In this work we introduce reinforcement learning techniques for solving lexicographic multi-objective problems. These are problems that involve multiple reward signals, and where the goal is to learn a policy that maximises the first reward signal, and subject to this constraint also maximises the second reward signal, and so on. We present a family of both action-value and policy gradient algorithms that can be used to solve such problems, and prove that they converge to policies that are lexicographically optimal. We evaluate the scalability and performance of these algorithms empirically, and demonstrate their applicability in practical settings. As a more specific application, we show how our algorithms can be used to impose safety constraints on the behaviour of an agent, and compare their performance in this context with that of other constrained reinforcement learning algorithms.

Lexicographic inference is a well-known and popular approach to reasoning with non-monotonic conditionals. It is a logic of very high-quality, as it extends rational closure and avoids the so-called drowning problem. It seems, however, this high quality comes at a cost, as reasoning on the basis of lexicographic inference is of high computational complexity. In this paper, we show that lexicographic inference satisfies syntax splitting, which means that we can restrict our attention to parts of the belief base that share atoms with a given query, thus seriously restricting the computational costs for many concrete queries. Furthermore, we make some observations on the relationship between c-representations and lexicographic inference, and reflect on the relation between syntax splitting and the drowning problem.

Deep learning models for different NLP tasks often rely on pre-trained word embeddings, that is, vector representations of words. Therefore, it is crucial to evaluate pre-trained word embeddings independently of downstream tasks. Such evaluations try to assess whether the geometry induced by a word embedding captures connections made in natural language, such as, analogies, clustering of words, or word similarities. Here, traditionally, similarity is measured by comparison to human judgment. However, explicitly annotating word pairs with similarity scores by surveying humans is expensive. We tackle this problem by formulating a similarity measure that is based on an agent for routing the Wikipedia hyperlink graph. In this graph, word similarities are implicitly encoded by edges between articles. We show on the English Wikipedia that our measure correlates well with a large group of traditional similarity measures, while covering a much larger proportion of words and avoiding explicit human labeling. Moreover, since Wikipedia is available in more than 300 languages, our measure can easily be adapted to other languages, in contrast to traditional similarity measures.