IJCAI 2022 论文列表

The demand for machine learning systems that can provide both transparency and fairness is constantly growing. Since the concept of fairness depends on the context, studies in the literature have proposed various formalisation and mitigation strategies. In this work, we propose a novel, flexible, discrimination-aware classifier that allows the user to: (i) select and mitigate the desired fairness criterion from a set of available options; (ii) implement more than one fairness criterion; (iii) handle more than one sensitive attribute; and (iv) specify the desired level of fairness to meet specific business needs or regulatory requirements. Our approach is based on an optimised extension to the decision-tree classifier, and aims to provide transparent and fair rules to the final users.

Decarbonization of global power systems significantly increases the operational uncertainty and modeling complexity that drive the necessity of widely exploiting cutting-edge Deep Reinforcement Learning (DRL) technologies to realize adaptive and real-time emergency control, which is the last resort for system stability and resiliency. The vulnerability of the DRL-based emergency control scheme may lead to severe real-world security issues if it can not be fully explored before implementing it practically. To this end, this is the first work that comprehensively investigates adversarial attacks and defense mechanisms for DRL-based power system emergency control. In particular, recovery-targeted (RT) adversarial attacks are designed for gradient-based approaches, aiming to dramatically degrade the effectiveness of the conducted emergency control actions to prevent the system from restoring to a stable state. Furthermore, the corresponding robust defense (RD) mechanisms are proposed to actively modify the observations based on the distances of sequential states. Experiments are conducted based on the standard IEEE reliability test system, and the results show that security risks indeed exist in the state-of-the-art DRL-based power system emergency control models. The effectiveness, stealthiness, instantaneity, and transferability of the proposed attacks and defense mechanisms are demonstrated with both white-box and black-box settings.

For problems requiring cooperation, many multiagent systems implement solutions among either individual agents or across an entire population towards a common goal. Multiagent teams are primarily studied when in conflict; however, organizational psychology (OP) highlights the benefits of teams among human populations for learning how to coordinate and cooperate. In this paper, we propose a new model of multiagent teams for reinforcement learning (RL) agents inspired by OP and early work on teams in artificial intelligence. We validate our model using complex social dilemmas that are popular in recent multiagent RL and find that agents divided into teams develop cooperative pro-social policies despite incentives to not cooperate. Furthermore, agents are better able to coordinate and learn emergent roles within their teams and achieve higher rewards compared to when the interests of all agents are aligned.

Deep convolutional neural networks (CNNs) have become dominant in the task of single image rain removal. Most of current CNN methods, however, suffer from the problem of overfitting on one single synthetic dataset as they neglect the intrinsic prior of the physical properties of rain streaks. To address this issue, we propose a simple but effective prior - Fourier prior to improve the generalization ability of an image rain removal model. The Fourier prior is a kind of property of rainy images. It is based on a key observation of us - replacing the Fourier amplitude of rainy images with that of clean images greatly suppresses the synthetic and real-world rain streaks. This means the amplitude contains most of the rain streak information and the phase keeps the similar structures of the background. So it is natural for single image rain removal to process the amplitude and phase information of the rainy images separately. In this paper, we develop a two-stage model where the first stage restores the amplitude of rainy images to clean rain streaks, and the second stage restores the phase information to refine fine-grained background structures. Extensive experiments on synthetic rainy data demonstrate the power of Fourier prior. Moreover, when trained on synthetic data, a robust generalization ability to real-world images can also be obtained. The code will be publicly available at https://github.com/willinglucky/ExploringFourier-Prior-for-Single-Image-Rain-Removal.

Partial label learning (PLL) is to learn a discriminative model under incomplete supervision, where each instance is annotated with a candidate label set. The basic principle of PLL is that the unknown correct label y of an instance x resides in its candidate label set s, i.e., P(y ∈ s | x) = 1. On which basis, current researches either directly model P(x | y) under different data generation assumptions or propose various surrogate multiclass losses, which all aim to encourage the model-based Pθ(y ∈ s | x)→1 implicitly. In this work, instead, we explicitly construct a binary classification task toward P(y ∈ s | x) based on the discriminative model, that is to predict whether the model-output label of x is one of its candidate labels. We formulate a novel risk estimator with estimation error bound for the proposed PLL binary classification risk. By applying logit adjustment based on disambiguation strategy, the practical approach directly maximizes Pθ(y ∈ s | x) while implicitly disambiguating the correct one from candidate labels simultaneously. Thorough experiments validate that the proposed approach achieves competitive performance against the state-of-the-art PLL methods.

Even though neural machine translation (NMT) has become the state-of-the-art solution for end-to-end translation, it still suffers from a lack of translation interpretability, which may be conveniently enhanced by explicit alignment learning (EAL), as performed in traditional statistical machine translation (SMT). To provide the benefits of both NMT and SMT, this paper presents a novel model design that enhances NMT with an additional training process for EAL, in addition to the end-to-end translation training. Thus, we propose two approaches an explicit alignment learning approach, in which we further remove the need for the additional alignment model, and perform embedding mixup with the alignment based on encoder--decoder attention weights in the NMT model. We conducted experiments on both small-scale (IWSLT14 De->En and IWSLT13 Fr->En) and large-scale (WMT14 En->De, En->Fr, WMT17 Zh->En) benchmarks. Evaluation results show that our EAL methods significantly outperformed strong baseline methods, which shows the effectiveness of EAL. Further explorations show that the translation improvements are due to a better spatial alignment of the source and target language embeddings. Our method improves translation performance without the need to increase model parameters and training data, which verifies that the idea of incorporating techniques of SMT into NMT is worthwhile.

Inconsistency-tolerant semantics have been proposed to provide meaningful query answers even in the presence of inconsistent knowledge. Recently, explainability has also become a prominent problem in different areas of AI. While the complexity of inconsistency-tolerant semantics is rather well-understood, not much attention has been paid yet to the problem of explaining query answers when inconsistencies may exist. Recent work on existential rules in the inconsistent setting has focused only on understanding why a query is entailed. In this paper, we address another important problem, which is explaining why a query is not entailed under an inconsistency-tolerant semantics. In particular, we consider three popular semantics, namely, the ABox repair, the intersection of repairs, and the intersection of closed repairs. We provide a thorough complexity analysis for a wide range of existential rule languages and for several complexity measures.

The aim of this work is to explain the observed behaviour of a hybrid system (HS). The explanation problem is cast as finding a trajectory of the HS that matches some observations. By using the formalism of hybrid automata (HA), we characterize the explanations as the language of a network of HA that comprises one automaton for the HS and another one for the observations, thus restricting the behaviour of the HS exclusively to trajectories that explain the observations. We observe that this problem corresponds to a reachability problem in model-checking, but that state-of-the-art model checkers struggle to find concrete trajectories. To overcome this issue we provide a formal mapping from HA to PDDL+ and show how to use an off-the-shelf automated planner. An experimental analysis over domains with piece-wise constant, linear and nonlinear dynamics reveals that the proposed PDDL+ approach is much more efficient than solving directly the explanation problem with model-checking solvers.

Recent work suggests to explain trade-offs between soft-goals in terms of their conflicts, i.e., minimal unsolvable soft-goal subsets. But this does not explain the conflicts themselves: Why can a given set of soft-goals not be jointly achieved? Here we approach that question in terms of the underlying constraints on plans in the task at hand, namely resource availability and time windows. In this context, a natural form of explanation for a soft-goal conflict is a minimal constraint relaxation under which the conflict disappears (``if the deadline was 1 hour later, it would work''). We explore algorithms for computing such explanations. A baseline is to simply loop over all relaxed tasks and compute the conflicts for each separately. We improve over this by two algorithms that leverage information -- conflicts, reachable states -- across relaxed tasks. We show that these algorithms can exponentially outperform the baseline in theory, and we run experiments confirming that advantage in practice.

In some preference aggregation scenarios, voters' preferences are highly structured: e.g., the set of candidates may have one-dimensional structure (so that voters' preferences are single-peaked) or be described by a binary decision tree (so that voters' preferences are group-separable). However, sometimes a single axis or a decision tree is insufficient to capture the voters' preferences; rather, there is a small number K of axes or decision trees such that each vote in the profile is consistent with one of these axes (resp., trees). In this work, we study the complexity of deciding whether voters' preferences can be explained in this manner. For K=2, we use the technique developed by Yang [2020, https://doi.org/10.3233/FAIA200099] in the context of single-peaked preferences to obtain a polynomial-time algorithm for several domains: value-restricted preferences, group-separable preferences, and a natural subdomain of group-separable preferences, namely, caterpillar group-separable preferences. For K > 2, the problem is known to be hard for single-peaked preferences; we establish that it is also hard for value-restricted and group-separable preferences. Our positive results for K=2 make use of forbidden minor characterizations of the respective domains; in particular, we establish that the domain of caterpillar group-separable preferences admits a forbidden minor characterization.

Responsible integration of deep neural networks into the design of trustworthy systems requires the ability to explain decisions made by these models. Explainability and transparency are critical for system analysis, certification, and human-machine teaming. We have recently demonstrated that neural stochastic differential equations (SDEs) present an explanation-friendly DNN architecture. In this paper, we present ExplainIt, an online tool for explaining AI decisions that uses neural SDEs to create visually sharper and more robust attributions than traditional residual neural networks. Our tool shows that the injection of noise in every layer of a residual network often leads to less noisy and less fragile integrated gradient attributions. The discrete neural stochastic differential equation model is trained on the ImageNet data set with a million images, and the demonstration produces robust attributions on images in the ImageNet validation library and on a variety of images in the wild. Our online tool is hosted publicly for educational purposes.

The existence of an anomaly detection method that is optimal for all domains is a myth. Thus, there exists a plethora of anomaly detection methods which increases every year for a wide variety of domains. But a strength can also be a weakness; given this massive library of methods, how can one select the best method for their application? Current literature is focused on creating new anomaly detection methods or large frameworks for experimenting with multiple methods at the same time. However, and especially as the literature continues to expand, an extensive evaluation of every anomaly detection method is simply not feasible. To reduce this evaluation burden, we present guidelines to intelligently choose the optimal anomaly detection methods based on the characteristics the time series displays such as seasonality, trend, level change concept drift, and missing time steps. We provide a comprehensive experimental validation and survey of twelve anomaly detection methods over different time series characteristics to form guidelines based on several metrics: the AUC (Area Under the Curve), windowed F-score, and Numenta Anomaly Benchmark (NAB) scoring model. Applying our methodologies can save time and effort by surfacing the most promising anomaly detection methods instead of experimenting extensively with a rapidly expanding library of anomaly detection methods, especially in an online setting.

Sum-Product Networks (SPNs) are expressive probabilistic models that provide exact, tractable inference. They achieve this efficiency by making use of local independence. On the other hand, mixtures of exchangeable variable models (MEVMs) are a class of tractable probabilistic models that make use of exchangeability of discrete random variables to render inference tractable. Exchangeability, which arises naturally in relational domains, has not been considered for efficient representation and inference in SPNs yet. The contribution of this paper is a novel probabilistic model which we call Exchangeability-Aware Sum-Product Networks (XSPNs). It contains both SPNs and MEVMs as special cases, and combines the ability of SPNs to efficiently learn deep probabilistic models with the ability of MEVMs to efficiently handle exchangeable random variables. We introduce a structure learning algorithm for XSPNs and empirically show that they can be more accurate than conventional SPNs when the data contains repeated, interchangeable parts.

Green Security Games have become a popular way to model scenarios involving the protection of natural resources, such as wildlife. Sensors (e.g. drones equipped with cameras) have also begun to play a role in these scenarios by providing real-time information. Incorporating both human and sensor defender resources strategically is the subject of recent work on Security Games with Signaling (SGS). However, current methods to solve SGS do not scale well in terms of time or memory. We therefore propose a novel approach to SGS, which, for the first time in this domain, employs an Evolutionary Computation paradigm: EASGS. EASGS effectively searches the huge SGS solution space via suitable solution encoding in a chromosome and a specially-designed set of operators. The operators include three types of mutations, each focusing on a particular aspect of the SGS solution, optimized crossover and a local coverage improvement scheme (a memetic aspect of EASGS). We also introduce a new set of benchmark games, based on dense or locally-dense graphs that reflect real-world SGS settings. In the majority of 342 test game instances, EASGS outperforms state-of-the-art methods, including a reinforcement learning method, in terms of time scalability, nearly constant memory utilization, and quality of the returned defender's strategies (expected payoffs).

When collaborating with an artificial intelligence (AI) system, we need to assess when to trust its recommendations. Suppose we mistakenly trust it in regions where it is likely to err. In that case, catastrophic failures may occur, hence the need for Bayesian approaches for reasoning and learning to determine the confidence (or epistemic uncertainty) in the probabilities of the queried outcome. Pure Bayesian methods, however, suffer from high computational costs. To overcome them, we revert to efficient and effective approximations. In this paper, we focus on techniques that take the name of evidential reasoning and learning from the process of Bayesian update of given hypotheses based on additional evidence. This paper provides the reader with a gentle introduction to the area of investigation, the up-to-date research outcomes, and the open questions still left unanswered.

Video deblurring with event information has attracted considerable attention. To help deblur each frame, existing methods usually compress a specific event sequence into a feature tensor with the same size as the corresponding video. However, this strategy neither considers the pixel-level spatial brightness changes nor the temporal correlation between events at each time step, resulting in insufficient use of spatio-temporal information. To address this issue, we propose a new Spatio-Temporal Relation-Attention network (STRA), for the specific event-based video deblurring. Concretely, to utilize spatial consistency between the frame and event, we model the brightness changes as an extra prior to aware blurring contexts in each frame; to record temporal relationship among different events, we develop a temporal memory block to restore long-range dependencies of event sequences continuously. In this way, the complementary information contained in the events and frames, as well as the correlation of neighboring events, can be fully utilized to recover spatial texture from events constantly. Experiments show that our STRA significantly outperforms several competing methods, e.g., on the HQF dataset, our network achieves up to 1.3 dB in terms of PSNR over the most advanced method. The code is available at https://github.com/Chengzhi-Cao/STRA.

Representation learning enables us to automatically extract generic feature representations from a dataset to solve another machine learning task. Recently, extracted feature representations by a representation learning algorithm and a simple predictor have exhibited state-of-the-art performance on several machine learning tasks. Despite its remarkable progress, there exist various ways to evaluate representation learning algorithms depending on the application because of the flexibility of representation learning. To understand the current applications of representation learning, we review evaluation methods of representation learning algorithms. On the basis of our evaluation survey, we also discuss the future direction of representation learning. The extended version, https://arxiv.org/abs/2204.08226, gives more detailed discussions and a survey on theoretical analyses.

Machine learning (ML) and artificial intelligence (AI) researchers play an important role in the ethics and governance of AI, including through their work, advocacy, and choice of employment. Nevertheless, this influential group's attitudes are not well understood, undermining our ability to discern consensuses or disagreements between AI/ML researchers. To examine these researchers' views, we conducted a survey of those who published in two top AI/ML conferences (N = 524). We compare these results with those from a 2016 survey of AI/ML researchers and a 2018 survey of the US public. We find that AI/ML researchers place high levels of trust in international organizations and scientific organizations to shape the development and use of AI in the public interest; moderate trust in most Western tech companies; and low trust in national militaries, Chinese tech companies, and Facebook. While the respondents were overwhelmingly opposed to AI/ML researchers working on lethal autonomous weapons, they are less opposed to researchers working on other military applications of AI, particularly logistics algorithms. A strong majority of respondents think that AI safety research should be prioritized more and a majority that ML institutions should conduct pre-publication review to assess potential harms. Being closer to the technology itself, AI/ML researchers are well placed to highlight new risks and develop technical solutions, so this novel data has broad relevance. The findings should help to improve how researchers, private sector executives, and policymakers think about regulations, governance frameworks, guiding principles, and national and international governance strategies for AI.

We study the relationship between the arrangement of neurons and the complexity of the ReLU-activated neural networks measured by the number of linear regions. More specifically, we provide both theoretical and empirical evidence for the point of view that shallow networks tend to have higher complexity than deep ones when the total number of neurons is fixed. In the theoretical part, we prove that this is the case for networks whose neurons in the hidden layers are arranged in the forms of 1x2n, 2xn and nx2; in the empirical part, we implement an algorithm that precisely tracks (hence counts) all the linear regions, and run it on networks with various structures. Although the time complexity of the algorithm is quite high, we verify that the problem of calculating the number of linear regions of a ReLU network is itself NP-hard. So currently there is no surprisingly efficient way to solve it. Roughly speaking, in the algorithm we divide the linear regions into subregions called the "activation regions", which are convex and easy to propagate through the network. The relationship between the number of the linear regions and that of the activation regions is also discussed.