IJCAI 2021 Papers

Nonlinear regression, although widely used in engineering, financial and security applications for automated decision making, is known to be vulnerable to training data poisoning. Targeted poisoning attacks may cause learning algorithms to fit decision functions with poor predictive performance. This paper presents a new analysis of local intrinsic dimensionality (LID) of nonlinear regression under such poisoning attacks within a Stackelberg game, leading to a practical defense. After adapting a gradient-based attack on linear regression that significantly impairs prediction capabilities to nonlinear settings, we consider a multi-step unsupervised black-box defense. The first step identifies samples that have the greatest influence on the learner's validation error; we then use the theory of local intrinsic dimensionality, which reveals the degree of being an outlier of data samples, to iteratively identify poisoned samples via a generative probabilistic model, and suppress their influence on the prediction function. Empirical validation demonstrates superior performance compared to a range of recent defenses.

Vision-and-Language (VL) pre-training has shown great potential on many related downstream tasks, such as Visual Question Answering (VQA), one of the most popular problems in the VL field. All of these pre-trained models (such as VisualBERT, ViLBERT, LXMERT and UNITER) are built with Transformer, which extends the classical attention mechanism to multiple layers and heads. To investigate why and how these models work on VQA so well, in this paper we explore the roles of individual heads and layers in Transformer models when handling 12 different types of questions. Specifically, we manually remove (chop) heads (or layers) from a pre-trained VisualBERT model at a time, and test it on different levels of questions to record its performance. As shown in the interesting echelon shape of the result matrices, experiments reveal different heads and layers are responsible for different question types, with higher-level layers activated by higher-level visual reasoning questions. Based on this observation, we design a dynamic chopping module that can automatically remove heads and layers of the VisualBERT at an instance level when dealing with different questions. Our dynamic chopping module can effectively reduce the parameters of the original model by 50%, while only damaging the accuracy by less than 1% on the VQA task.

Choosing a suitable algorithm from the myriads of different search heuristics is difficult when faced with a novel optimization problem. In this work, we argue that the purely academic question of what could be the best possible algorithm in a certain broad class of black-box optimizers can give fruitful indications in which direction to search for good established optimization heuristics. We demonstrate this approach on the recently proposed DLB benchmark, for which the only known results are O(n^3) runtimes for several classic evolutionary algorithms and an O(n^2 log n) runtime for an estimation-of-distribution algorithm. Our finding that the unary unbiased black-box complexity is only O(n^2) suggests the Metropolis algorithm as an interesting candidate and we prove that it solves the DLB problem in quadratic time. Since we also prove that better runtimes cannot be obtained in the class of unary unbiased algorithms, we shift our attention to algorithms that use the information of more parents to generate new solutions. An artificial algorithm of this type having an O(n log n) runtime leads to the result that the significance-based compact genetic algorithm (sig-cGA) can solve the DLB problem also in time O(n log n). Our experiments show a remarkably good performance of the Metropolis algorithm, clearly the best of all algorithms regarded for reasonable problem sizes.

Qualitative Choice Logic (QCL) and Conjunctive Choice Logic (CCL) are formalisms for preference handling, with especially QCL being well established in the field of AI. So far, analyses of these logics need to be done on a case-by-case basis, albeit they share several common features. This calls for a more general choice logic framework, with QCL and CCL as well as some of their derivatives being particular instantiations. We provide such a framework, which allows us, on the one hand, to easily define new choice logics and, on the other hand, to examine properties of different choice logics in a uniform setting. In particular, we investigate strong equivalence, a core concept in non-classical logics for understanding formula simplification, and computational complexity. Our analysis also yields new results for QCL and CCL. For example, we show that the main reasoning task regarding preferred models is ϴ₂P-complete for QCL and CCL, while being Δ₂P-complete for a newly introduced choice logic.

This paper proposes Characteristic Examples for effectively fingerprinting deep neural networks, featuring high-robustness to the base model against model pruning as well as low-transferability to unassociated models. This is the first work taking both robustness and transferability into consideration for generating realistic fingerprints, whereas current methods lack practical assumptions and may incur large false positive rates. To achieve better trade-off between robustness and transferability, we propose three kinds of characteristic examples: vanilla C-examples, RC-examples, and LTRC-example, to derive fingerprints from the original base model. To fairly characterize the trade-off between robustness and transferability, we propose Uniqueness Score, a comprehensive metric that measures the difference between robustness and transferability, which also serves as an indicator to the false alarm problem. Extensive experiments demonstrate that the proposed characteristic examples can achieve superior performance when compared with existing fingerprinting methods. In particular, for VGG ImageNet models, using LTRC-examples gives 4X higher uniqueness score than the baseline method and does not incur any false positives.

Incorporating humans into AI planning is an important feature of flexible planning technology. Such human integration allows to incorporate previously unknown constraints, and is also an integral part of automated modeling assistance. As a foundation for integrating user requests, we study the computational complexity of determining the existence of changes to an existing model, such that the resulting model allows for specific user-provided solutions. We are provided with a planning problem modeled either in the classical (non-hierarchical) or hierarchical task network (HTN) planning formalism, as well as with a supposed-to-be solution plan, which is actually not a solution for the current model. Considering changing decomposition methods as well as preconditions and effects of actions, we show that most change requests are NP-complete though some turn out to be tractable.

Deep learning is revolutionizing predictive healthcare, including recommending medications to patients with complex health conditions. Existing approaches focus on predicting all medications for the current visit, which often overlaps with medications from previous visits. A more clinically relevant task is to identify medication changes. In this paper, we propose a new recurrent residual networks, named MICRON, for medication change prediction. MICRON takes the changes in patient health records as input and learns to update a hid- den medication vector and the medication set recurrently with a reconstruction design. The medication vector is like the memory cell that encodes longitudinal information of medications. Unlike traditional methods that require the entire patient history for prediction, MICRON has a residual-based inference that allows for sequential updating based only on new patient features (e.g., new diagnoses in the recent visit), which is efficient. We evaluated MICRON on real inpatient and outpatient datasets. MICRON achieves 3.5% and 7.8% relative improvements over the best baseline in F1 score, respectively. MICRON also requires fewer parameters, which significantly reduces the training time to 38.3s per epoch with 1.5× speed-up.

In the last few years, Gradient Boosting Decision Trees (GBDTs) have been widely used in various applications such as online advertising and spam filtering. However, GBDT training is often a key performance bottleneck for such data science pipelines, especially for training a large number of deep trees on large data sets. Thus, many parallel and distributed GBDT systems have been researched and developed to accelerate the training process. In this survey paper, we review the recent GBDT systems with respect to accelerations with emerging hardware as well as cluster computing, and compare the advantages and disadvantages of the existing implementations. Finally, we present the research opportunities and challenges in designing fast next generation GBDT systems.

There have been increasing concerns about Artificial Intelligence (AI) due to its unfathomable potential power. To make AI address ethical challenges and shun undesirable outcomes, researchers proposed to develop socially responsible AI (SRAI). One of these approaches is causal learning (CL). We survey state-of-the-art methods of CL for SRAI. We begin by examining the seven CL tools to enhance the social responsibility of AI, then review how existing works have succeeded using these tools to tackle issues in developing SRAI such as fairness. The goal of this survey is to bring forefront the potentials and promises of CL for SRAI.

Discovering causal structures among latent factors from observed data is a particularly challenging problem. Despite some efforts for this problem, existing methods focus on the single-domain data only. In this paper, we propose Multi-Domain Linear Non-Gaussian Acyclic Models for LAtent Factors (MD-LiNA), where the causal structure among latent factors of interest is shared for all domains, and we provide its identification results. The model enriches the causal representation for multi-domain data. We propose an integrated two-phase algorithm to estimate the model. In particular, we first locate the latent factors and estimate the factor loading matrix. Then to uncover the causal structure among shared latent factors of interest, we derive a score function based on the characterization of independence relations between external influences and the dependence relations between multi-domain latent factors and latent factors of interest. We show that the proposed method provides locally consistent estimators. Experimental results on both synthetic and real-world data demonstrate the efficacy and robustness of our approach.

Ontology-mediated query answering (OMQA) employs structured knowledge and automated reasoning in order to facilitate access to incomplete and possibly heterogeneous data. While most research on OMQA adopts (unions of) conjunctive queries as the query language, there has been recent interest in handling queries that involve counting. In this paper, we advance this line of research by investigating cardinality queries (which correspond to Boolean atomic counting queries) coupled with DL-Lite ontologies. Despite its apparent simplicity, we show that such an OMQA setting gives rise to rich and complex behaviour. While we prove that cardinality query answering is tractable (TC0) in data complexity when the ontology is formulated in DL-Lite-core, the problem becomes coNP-hard as soon as role inclusions are allowed. For DL-Lite-pos-H (which allows only positive axioms), we establish a P-coNP dichotomy and pinpoint the TC0 cases; for DL-Lite-core-H (allowing also negative axioms), we identify new sources of coNP complexity and also exhibit L-complete cases. Interestingly, and in contrast to related tractability results, we observe that the canonical model may not give the optimal count value in the tractable cases, which led us to develop an entirely new approach based upon exploring a space of strategies to determine the minimum possible number of query matches.

Molecular interactions are significant resources for analyzing sophisticated biological systems. Identification of multifarious molecular interactions attracts increasing attention in biomedicine, bioinformatics, and human healthcare communities. Recently, a plethora of methods have been proposed to reveal molecular interactions in one specific domain. However, existing methods heavily rely on features or structures involving molecules, which limits the capacity of transferring the models to other tasks. Therefore, generalized models for the multifarious molecular interaction prediction (MIP) are in demand. In this paper, we propose a contrastive self-supervised graph neural network (CSGNN) to predict molecular interactions. CSGNN injects a mix-hop neighborhood aggregator into a graph neural network (GNN) to capture high-order dependency in the molecular interaction networks and leverages a contrastive self-supervised learning task as a regularizer within a multi-task learning paradigm to enhance the generalization ability. Experiments on seven molecular interaction networks show that CSGNN outperforms classic and state-of-the-art models. Comprehensive experiments indicate that the mix-hop aggregator and the self-supervised regularizer can effectively facilitate the link inference in multifarious molecular networks.

Recently, hashing is widely used in approximate nearest neighbor search for its storage and computational efficiency. Most of the unsupervised hashing methods learn to map images into semantic similarity-preserving hash codes by constructing local semantic similarity structure from the pre-trained model as the guiding information, i.e., treating each point pair similar if their distance is small in feature space. However, due to the inefficient representation ability of the pre-trained model, many false positives and negatives in local semantic similarity will be introduced and lead to error propagation during the hash code learning. Moreover, few of the methods consider the robustness of models, which will cause instability of hash codes to disturbance. In this paper, we propose a new method named Comprehensive sImilarity Mining and cOnsistency learNing (CIMON). First, we use global refinement and similarity statistical distribution to obtain reliable and smooth guidance. Second, both semantic and contrastive consistency learning are introduced to derive both disturb-invariant and discriminative hash codes. Extensive experiments on several benchmark datasets show that the proposed method outperforms a wide range of state-of-the-art methods in both retrieval performance and robustness.

In many real-world scenarios, a team of agents must coordinate with each other to compete against an opponent. The challenge of solving this type of game is that the team's joint action space grows exponentially with the number of agents, which results in the inefficiency of the existing algorithms, e.g., Counterfactual Regret Minimization (CFR). To address this problem, we propose a new framework of CFR: CFR-MIX. Firstly, we propose a new strategy representation that represents a joint action strategy using individual strategies of all agents and a consistency relationship to maintain the cooperation between agents. To compute the equilibrium with individual strategies under the CFR framework, we transform the consistency relationship between strategies to the consistency relationship between the cumulative regret values. Furthermore, we propose a novel decomposition method over cumulative regret values to guarantee the consistency relationship between the cumulative regret values. Finally, we introduce our new algorithm CFR-MIX which employs a mixing layer to estimate cumulative regret values of joint actions as a non-linear combination of cumulative regret values of individual actions. Experimental results show that CFR-MIX outperforms existing algorithms on various games significantly.

Affordances describe the possibilities for an agent to perform actions with an object. While the significance of the affordance concept has been previously studied from varied perspectives, such as psychology and cognitive science, these approaches are not always sufficient to enable direct transfer, in the sense of implementations, to artificial intelligence (AI)-based systems and robotics. However, many efforts have been made to pragmatically employ the concept of affordances, as it represents great potential for AI agents to effectively bridge perception to action. In this survey, we review and find common ground amongst different strategies that use the concept of affordances within robotic tasks, and build on these methods to provide guidance for including affordances as a mechanism to improve autonomy. To this end, we outline common design choices for building representations of affordance relations, and their implications on the generalisation capabilities of an agent when facing previously unseen scenarios. Finally, we identify and discuss a range of interesting research directions involving affordances that have the potential to improve the capabilities of an AI agent.

In this paper, we consider the problem of designing budget-feasible mechanisms for selecting agents with private costs from various groups to ensure proportional representation, where the minimum proportion of the selected agents from each group is maximized. Depending on agents' membership in the groups, we consider two main models: single group setting where each agent belongs to only one group, and multiple group setting where each agent may belong to multiple groups. We propose novel budget-feasible proportion-representative mechanisms for these models, which can select representative agents from different groups. The proposed mechanisms guarantee theoretical properties of individual rationality, budget-feasibility, truthfulness, and approximation performance on proportional representation.

The Nash social welfare (NSW) is a well-known social welfare measurement that balances individual utilities and the overall efficiency. In the context of fair allocation of indivisible goods, it has been shown by Caragiannis et al. (EC 2016 and TEAC 2019) that an allocation maximizing the NSW is envy-free up to one good (EF1). In this paper, we are interested in the fairness of the NSW in a budget-feasible allocation problem, in which each item has a cost that will be incurred to the agent it is allocated to, and each agent has a budget constraint on the total cost of items she receives. We show that a budget-feasible allocation that maximizes the NSW achieves a 1/4-approximation of EF1 and the approximation ratio is tight. The approximation ratio improves gracefully when the items have small costs compared with the agents' budgets; it converges to 1/2 when the budget-cost ratio approaches infinity.

Discounting future costs and rewards is a common practice in accounting, game theory, and machine learning. In spite of this, existing logics for reasoning about strategies with cost and resource constraints do not account for discounting. The paper proposes a sound and complete logical system for reasoning about budget-constrained strategic abilities that incorporates discounting into its semantics.

Many planning problems are too hard to solve optimally. In bounded-cost search, one attempts to find, as quickly as possible, a plan that costs no more than a user-provided absolute cost bound. Several algorithms have been previously proposed for this setting, including Potential Search (PTS) and Bounded-cost Explicit Estimation Search (BEES). BEES attempts to improve on PTS by predicting whether nodes will lead to plans within the cost bound or not. This paper introduces a relatively simple algorithm, Expected Effort Search (XES), which uses not just point estimates but belief distributions in order to estimate the probability that a node will lead to a plan within the bound. XES's expansion order minimizes expected search time in a simplified formal model. Experimental results on standard planning and search benchmarks show that it consistently exhibits strong performance, outperforming both PTS and BEES. We also derive improved variants of BEES that can exploit belief distributions. These new methods advance the recent trend of taking advantage of uncertainty estimates in deterministic single-agent search.