IJCAI 2021 Papers

Differentially private machine learning trains models while protecting privacy of the sensitive training data. The key to obtain differentially private models is to introduce noise/randomness to the training process. In particular, existing differentially private machine learning methods add noise to the training data, the gradients, the loss function, and/or the model itself. Bagging, a popular ensemble learning framework, randomly creates some subsamples of the training data, trains a base model for each subsample using a base learner, and takes majority vote among the base models when making predictions. Bagging has intrinsic randomness in the training process as it randomly creates subsamples. Our major theoretical results show that such intrinsic randomness already makes Bagging differentially private without the needs of additional noise. Moreover, we prove that if no assumptions about the base learner are made, our derived privacy guarantees are tight. We empirically evaluate Bagging on MNIST and CIFAR10. Our experimental results demonstrate that Bagging achieves significantly higher accuracies than state-of-the-art differentially private machine learning methods with the same privacy budgets.

The compositional minimax problem covers plenty of machine learning models such as the distributionally robust compositional optimization problem. However, it is yet another understudied problem to optimize the compositional minimax problem. In this paper, we develop a novel efficient stochastic compositional gradient descent ascent method for optimizing the compositional minimax problem. Moreover, we establish the theoretical convergence rate of our proposed method. To the best of our knowledge, this is the first work achieving such a convergence rate for the compositional minimax problem. Finally, we conduct extensive experiments to demonstrate the effectiveness of our proposed method.

We study the secretary problem in multi-agent environments. In the standard secretary problem, a sequence of arbitrary awards arrive online, in a random order, and a single decision maker makes an immediate and irrevocable decision whether to accept each award upon its arrival. The requirement to make immediate decisions arises in many cases due to an implicit assumption regarding competition. Namely, if the decision maker does not take the offered award immediately, it will be taken by someone else. We introduce a novel multi-agent secretary model, in which the competition is explicit. In our model, multiple agents compete over the arriving awards, but the decisions need not be immediate; instead, agents may select previous awards as long as they are available (i.e., not taken by another agent). If an award is selected by multiple agents, ties are broken either randomly or according to a global ranking. This induces a multi-agent game in which the time of selection is not enforced by the rules of the games, rather it is an important component of the agent's strategy. We study the structure and performance of equilibria in this game. For random tie breaking, we characterize the equilibria of the game, and show that the expected social welfare in equilibrium is nearly optimal, despite competition among the agents. For ranked tie breaking, we give a full characterization of equilibria in the 3-agent game, and show that as the number of agents grows, the winning probability of every agent under non-immediate selections approaches her winning probability under immediate selections.

Stubborn sets are a pruning technique for state-space search which is well established in optimal classical planning. In this paper, we show that weak stubborn sets introduced in recent work in planning are actually not weak stubborn sets in Valmari's original sense. Based on this finding, we introduce weak stubborn sets in the original sense for planning by providing a generalized definition analogously to generalized strong stubborn sets in previous work. We discuss the relationship of strong, weak and the previously called weak stubborn sets, thus providing a further step in getting an overall picture of the stubborn set approach in planning.

Several methods have recently been developed for computing attributions of a neural network's prediction over the input features. However, these existing approaches for computing attributions are noisy and not robust to small perturbations of the input. This paper uses the recently identified connection between dynamical systems and residual neural networks to show that the attributions computed over neural stochastic differential equations (SDEs) are less noisy, visually sharper, and quantitatively more robust. Using dynamical systems theory, we theoretically analyze the robustness of these attributions. We also experimentally demonstrate the efficacy of our approach in providing smoother, visually sharper and quantitatively robust attributions by computing attributions for ImageNet images using ResNet-50, WideResNet-101 models and ResNeXt-101 models.

Recently, the teacher-student knowledge distillation framework has demonstrated its potential in training Graph Neural Networks (GNNs). However, due to the difficulty of training over-parameterized GNN models, one may not easily obtain a satisfactory teacher model for distillation. Furthermore, the inefficient training process of teacher-student knowledge distillation also impedes its applications in GNN models. In this paper, we propose the first teacher-free knowledge distillation method for GNNs, termed GNN Self-Distillation (GNN-SD), that serves as a drop-in replacement of the standard training process. The method is built upon the proposed neighborhood discrepancy rate (NDR), which quantifies the non-smoothness of the embedded graph in an efficient way. Based on this metric, we propose the adaptive discrepancy retaining (ADR) regularizer to empower the transferability of knowledge that maintains high neighborhood discrepancy across GNN layers. We also summarize a generic GNN-SD framework that could be exploited to induce other distillation strategies. Experiments further prove the effectiveness and generalization of our approach, as it brings: 1) state-of-the-art GNN distillation performance with less training cost, 2) consistent and considerable performance enhancement for various popular backbones.

Recommender systems personalize content by recommending items to users. Item recommendation algorithms are evaluated by metrics that compare the positions of truly relevant items among the recommended items. To speed up the computation of metrics, recent work often uses sampled metrics where only a smaller set of random items and the relevant items are ranked. This paper investigates such sampled metrics and shows that they are inconsistent with their exact counterpart, in the sense that they do not persist relative statements, e.g., recommender A is better than B, not even in expectation. We show that it is possible to improve the quality of the sampled metrics by applying a correction. We conclude with an empirical evaluation of the naive sampled metrics and their corrected variants. Our work suggests that sampling should be avoided for metric calculation, however if an experimental study needs to sample, the proposed corrections can improve the estimates.

Learning from unordered sets is a fundamental learning setup, recently attracting increasing attention. Research in this area has focused on the case where elements of the set are represented by feature vectors, and far less emphasis has been given to the common case where set elements themselves adhere to their own symmetries. That case is relevant to numerous applications, from deblurring image bursts to multi-view 3D shape recognition and reconstruction. In this paper, we present a principled approach to learning sets of general symmetric elements. We first characterize the space of linear layers that are equivariant both to element reordering and to the inherent symmetries of elements, like translation in the case of images. We further show that networks that are composed of these layers, called Deep Sets for Symmetric Elements layers (DSS), are universal approximators of both invariant and equivariant functions, and that these networks are strictly more expressive than Siamese networks. DSS layers are also straightforward to implement. Finally, we show that they improve over existing set-learning architectures in a series of experiments with images, graphs, and point clouds.

We build on abduction-based explanations for machine learning and develop a method for computing local explanations for neural network models in natural language processing (NLP). Our explanations comprise a subset of the words of the input text that satisfies two key features: optimality w.r.t. a user-defined cost function, such as the length of explanation, and robustness, in that they ensure prediction invariance for any bounded perturbation in the embedding space of the left-out words. We present two solution algorithms, respectively based on implicit hitting sets and maximum universal subsets, introducing a number of algorithmic improvements to speed up convergence of hard instances. We show how our method can be configured with different perturbation sets in the embedded space and used to detect bias in predictions by enforcing include/exclude constraints on biased terms, as well as to enhance existing heuristic-based NLP explanation frameworks such as Anchors. We evaluate our framework on three widely used sentiment analysis tasks and texts of up to 100 words from SST, Twitter and IMDB datasets, demonstrating the effectiveness of the derived explanations.

Random Forest (RFs) are among the most widely used Machine Learning (ML) classifiers. Even though RFs are not interpretable, there are no dedicated non-heuristic approaches for computing explanations of RFs. Moreover, there is recent work on polynomial algorithms for explaining ML models, including naive Bayes classifiers. Hence, one question is whether finding explanations of RFs can be solved in polynomial time. This paper answers this question negatively, by proving that computing one PI-explanation of an RF is D^P-hard. Furthermore, the paper proposes a propositional encoding for computing explanations of RFs, thus enabling finding PI-explanations with a SAT solver. This contrasts with earlier work on explaining boosted trees (BTs) and neural networks (NNs), which requires encodings based on SMT/MILP. Experimental results, obtained on a wide range of publicly available datasets, demonstrate that the proposed SAT-based approach scales to RFs of sizes common in practical applications. Perhaps more importantly, the experimental results demonstrate that, for the vast majority of examples considered, the SAT-based approach proposed in this paper significantly outperforms existing heuristic approaches.

Abstract argumentation as defined by Dung in his seminal 1995 paper is by now a major research area in knowledge representation and reasoning. Dynamics of abstract argumentation frameworks (AFs) as well as syntactical consequences of semantical facts of them are the central issues of this paper. The first main part is engaged with the systematical study of the influence of attackers and supporters regarding the acceptability status of whole sets and/or single arguments. In particular, we investigate the impact of addition or removal of arguments, a line of research that has been around for more than a decade. Apart from entirely new results, we revisit, generalize and sum up similar results from the literature. To gain a comprehensive formal and intuitive understanding of the behavior of AFs we put special effort in comparing different kind of semantics. We concentrate on classical admissibility-based semantics and also give pointers to semantics based on naivity and weak admissibility, a recently introduced mediating approach. In the second main part we show how to infer syntactical information from semantical one. For instance, it is well-known that if a finite AF possesses no stable extension, then it has to contain an odd-cycle. In this paper, we even present a characterization of this issue. Moreover, we show that the change of the number of extensions if adding or removing an argument allows to conclude the existence of certain even or odd cycles in the considered AF without having further information.

An important issue in ML consists in developing approaches exploiting background knowledge T for improving the accuracy and the robustness of learned classifiers C. Delegating the classification task to a Boolean circuit Σ exhibiting the same input-output behaviour as C, the problem of exploiting T within C can be viewed as a belief change scenario. However, usual change operations are not suited to the task of modifying the classifier encoding Σ in a minimal way, to make it complying with T. To fill the gap, we present a new belief change operation, called rectification. We characterize the family of rectification operators from an axiomatic perspective and exhibit operators from this family. We identify the standard belief change postulates that every rectification operator satisfies and those it does not. We also focus on some computational aspects of rectification and compliance.

Traffic simulators act as an essential component in the operating and planning of transportation systems. Conventional traffic simulators usually employ a calibrated physical car-following model to describe vehicles' behaviors and their interactions with traffic environment. However, there is no universal physical model that can accurately predict the pattern of vehicle's behaviors in different situations. A fixed physical model tends to be less effective in a complicated environment given the non-stationary nature of traffic dynamics. In this paper, we formulate traffic simulation as an inverse reinforcement learning problem, and propose a parameter sharing adversarial inverse reinforcement learning model for dynamics-robust simulation learning. Our proposed model is able to imitate a vehicle's trajectories in the real world while simultaneously recovering the reward function that reveals the vehicle's true objective which is invariant to different dynamics. Extensive experiments on synthetic and real-world datasets show the superior performance of our approach compared to state-of-the-art methods and its robustness to variant dynamics of traffic.

Deep neural networks have been shown to be very powerful tools for object detection in various scenes. Their remarkable performance, however, heavily depends on the availability of a large number of high quality labeled data, which are time-consuming and costly to acquire for scenes with densely packed objects. We present a novel semi-supervised approach to addressing this problem, which is designed based on a common teacher-student model, integrated with a novel intersection-over-union (IoU) aware consistency loss and a new proposal consistency loss. The IoU-aware consistency loss evaluates the IoU over the prediction pairs of the teacher model and the student model, which enforces the prediction of the student model to approach closely to that of the teacher model. The IoU-aware consistency loss also reweights the importance of different prediction pairs to suppress the low-confident pairs. The proposal consistency loss ensures proposal consistency between the two models, making it possible to involve the region proposal network in the training process with unlabeled data. We also construct a new dataset, namely RebarDSC, containing 2,125 rebar images annotated with 350,348 bounding boxes in total (164.9 annotations per image average), to evaluate the proposed method. Extensive experiments are conducted over both the RebarDSC dataset and the famous large public dataset SKU-110K. Experimental results corroborate that the proposed method is able to improve the object detection performance in densely packed scenes, consistently outperforming state-of-the-art approaches. Dataset is available in https://github.com/Armin1337/RebarDSC.

Novelty detection is the process of determining whether a query example differs from the learned training distribution. Previous generative adversarial networks based methods and self-supervised approaches suffer from instability training, mode dropping, and low discriminative ability. We overcome such problems by introducing a novel decoder-encoder framework. Firstly, a generative network (decoder) learns the representation by mapping the initialized latent vector to an image. In particular, this vector is initialized by considering the entire distribution of training data to avoid the problem of mode-dropping. Secondly, a contrastive network (encoder) aims to ``learn to compare'' through mutual information estimation, which directly helps the generative network to obtain a more discriminative representation by using a negative data augmentation strategy. Extensive experiments show that our model has significant superiority over cutting-edge novelty detectors and achieves new state-of-the-art results on various novelty detection benchmarks, e.g. CIFAR10 and DCASE. Moreover, our model is more stable for training in a non-adversarial manner, compared to other adversarial based novelty detection methods.

Zero-shot sketch-based image retrieval (ZS-SBIR), which aims to retrieve photos with sketches under the zero-shot scenario, has shown extraordinary talents in real-world applications. Most existing methods leverage language models to generate class-prototypes and use them to arrange the locations of all categories in the common space for photos and sketches. Although great progress has been made, few of them consider whether such pre-defined prototypes are necessary for ZS-SBIR, where locations of unseen class samples in the embedding space are actually determined by visual appearance and a visual embedding actually performs better. To this end, we propose a novel Norm-guided Adaptive Visual Embedding (NAVE) model, for adaptively building the common space based on visual similarity instead of language-based pre-defined prototypes. To further enhance the representation quality of unseen classes for both photo and sketch modality, modality norm discrepancy and noisy label regularizer are jointly employed to measure and repair the modality bias of the learned common embedding. Experiments on two challenging datasets demonstrate the superiority of our NAVE over state-of-the-art competitors.

Although distributional reinforcement learning (DRL) has been widely examined in the past few years, there are two open questions people are still trying to address. One is how to ensure the validity of the learned quantile function, the other is how to efficiently utilize the distribution information. This paper attempts to provide some new perspectives to encourage the future in-depth studies in these two fields. We first propose a non-decreasing quantile function network (NDQFN) to guarantee the monotonicity of the obtained quantile estimates and then design a general exploration framework called distributional prediction error (DPE) for DRL which utilizes the entire distribution of the quantile function. In this paper, we not only discuss the theoretical necessity of our method but also show the performance gain it achieves in practice by comparing with some competitors on Atari 2600 Games especially in some hard-explored games.

Automatic pain recognition is paramount for medical diagnosis and treatment. The existing works fall into three categories: assessing facial appearance changes, exploiting physiological cues, or fusing them in a multi-modal manner. However, (1) appearance changes are easily affected by subjective factors which impedes objective pain recognition. Besides, the appearance-based approaches ignore long-range spatial-temporal dependencies that are important for modeling expressions over time; (2) the physiological cues are obtained by attaching sensors on human body, which is inconvenient and uncomfortable. In this paper, we present a novel multi-task learning framework which encodes both appearance changes and physiological cues in a non-contact manner for pain recognition. The framework is able to capture both local and long-range dependencies via the proposed attention mechanism for the learned appearance representations, which are further enriched by temporally attended physiological cues (remote photoplethysmography, rPPG) that are recovered from videos in the auxiliary task. This framework is dubbed rPPG-enriched Spatio-Temporal Attention Network (rSTAN) and allows us to establish the state-of-the-art performance of non-contact pain recognition on publicly available pain databases. It demonstrates that rPPG predictions can be used as an auxiliary task to facilitate non-contact automatic pain recognition.

We consider a problem wherein jobs arrive at random times and assume random values. Upon each job arrival, the decision-maker must decide immediately whether or not to accept the job and gain the value on offer as a reward, with the constraint that they may only accept at most n jobs over some reference time period. The decision-maker only has access to M independent realisations of the job arrival process. We propose an algorithm, Non-Parametric Sequential Allocation (NPSA), for solving this problem. Moreover, we prove that the expected reward returned by the NPSA algorithm converges in probability to optimality as M grows large. We demonstrate the effectiveness of the algorithm empirically on synthetic data and on public fraud-detection datasets, from where the motivation for this work is derived.