IJCAI 2023 Papers

AllDifferent constraint is widely used in Constraint Programming to model real world problems. Existing Generalized Arc Consistency (GAC) algorithms map an AllDifferent constraint onto a bipartite graph and utilize the structure of Strongly Connected Components (SCCs) in the graph to filter values. Calculating SCCs is time-consuming in the existing algorithms, so we propose a novel GAC algorithm for AllDifferent constraint in this paper, which eliminates the computation of SCCs. We prove that all redundant edges in the bipartite graph point to some alternating cycles. Our algorithm exploits this property and uses a more efficient method to filter values, which is based on breadth-first search. Experimental results on the XCSP3 benchmark suite show that our algorithm considerably outperforms the state-of-the-art GAC algorithms.

Vaccines have proven to be extremely effective in preventing the spread of COVID-19 and potentially ending the pandemic. Lack of access caused many people not getting vaccinated early, so states such as Virginia deployed mobile vaccination sites in order to distribute vaccines across the state. Here we study the problem of deciding where these facilities should be placed and moved over time in order to minimize the distance each person needs to travel in order to be vaccinated. Traditional facility location models for this problem fail to incorporate the fact that our facilities are mobile (i.e., they can move over time). To this end, we instead model vaccine distribution as the Dynamic k-Supplier problem and give the first approximation algorithms for this problem. We then run extensive simulations on real world datasets to show the efficacy of our methods. In particular, we find that natural baselines for Dynamic k-Supplier cannot take advantage of the mobility of the facilities, and perform worse than non-mobile k-Supplier algorithms.

Most existing studies on Sign Language Translation (SLT) employ AutoRegressive Decoding Mechanism (AR-DM) to generate target sentences. However, the main disadvantage of the AR-DM is high inference latency. To address this problem, we introduce Non-AutoRegressive Decoding Mechanism (NAR-DM) into SLT, which generates the whole sentence at once. Meanwhile, to improve its decoding ability, we integrate the advantages of curriculum learning and NAR-DM and propose a Curriculum-based NAR Decoder (CND). Specifically, the lower layers of the CND are expected to predict simple tokens that could be predicted correctly using source-side information solely. Meanwhile, the upper layers could predict complex tokens based on the lower layers' predictions. Therefore, our CND significantly reduces the model's inference latency while maintaining its competitive performance. Moreover, to further boost the performance of our CND, we propose a mutual learning framework, containing two decoders, i.e., an AR decoder and our CND. We jointly train the two decoders and minimize the KL divergence between their outputs, which enables our CND to learn the forward sequential knowledge from the strengthened AR decoder. Experimental results on PHOENIX2014T and CSL-Daily demonstrate that our model consistently outperforms all competitive baselines and achieves 7.92/8.02× speed-up compared to the AR SLT model respectively. Our source code is available at https://github.com/yp20000921/CND.

Constructing decision trees online is a classical machine learning problem. Existing works often assume that features are readily available for each incoming data point. However, in many real world applications, both feature values and the labels are unknown a priori and can only be obtained at a cost. For example, in medical diagnosis, doctors have to choose which tests to perform (i.e., making costly feature queries) on a patient in order to make a diagnosis decision (i.e., predicting labels). We provide a fresh perspective to tackle this practical challenge. Our framework consists of an active planning oracle embedded in an online learning scheme for which we investigate several information acquisition functions. Specifically, we employ a surrogate information acquisition function based on adaptive submodularity to actively query feature values with a minimal cost, while using a posterior sampling scheme to maintain a low regret for online prediction. We demonstrate the efficiency and effectiveness of our framework via extensive experiments on various real-world datasets. Our framework also naturally adapts to the challenging setting of online learning with concept drift and is shown to be competitive with baseline models while being more flexible.

Video games feature a dynamic environment where locations of objects (e.g., characters, equipment, weapons, vehicles etc.) frequently change within the game world. Although searching for relevant nearby objects in such a dynamic setting is a fundamental operation, this problem has received little research attention. In this paper, we propose a simple lightweight index, called Grid Tree, to store objects and their associated textual data. Our index can be efficiently updated with the underlying updates such as object movements, and supports a variety of object search queries, including k nearest neighbors (returning the k closest objects), keyword k nearest neighbors (returning the k closest objects that satisfy query keywords), and several other variants. Our extensive experimental study, conducted on standard game maps benchmarks and real-world keywords, demonstrates that our approach has up to 2 orders of magnitude faster update times for moving objects compared to state-of-the-art approaches such as navigation mesh and IR-tree. At the same time, query performance of our approach is similar to or better than that of IR-tree and up to two orders of magnitude faster than the other competitor.

As model finetuning is central to the modern NLP, we set to maximize its efficiency. Motivated by redundancy in training examples and the sheer sizes of pretrained models, we exploit a key opportunity: training only on important data. To this end, we set to filter training examples in a streaming fashion, in tandem with training the target model. Our key techniques are two: (1) automatically determine a training loss threshold for skipping backward training passes; (2) run a meta predictor for further skipping forward training passes. We integrate the above techniques in a holistic, three-stage training pro- cess. On a diverse set of benchmarks, our method reduces the required training examples by up to 5.3× and training time by up to 6.8×, while only seeing minor accuracy degradation. Our method is effective even for training one epoch, where each training example is encountered only once. It is simple to implement and is compatible with the existing finetuning techniques. Code is available at: https://github.com/xo28/efficient-NLP-multistage-training

Recovering the shape and appearance of real-world objects from natural 2D images is a long-standing and challenging inverse rendering problem. In this paper, we introduce a novel hybrid differentiable rendering method to efficiently reconstruct the 3D geometry and reflectance of a scene from multi-view images captured by conventional hand-held cameras. Our method follows an analysis-by-synthesis approach and consists of two phases. In the initialization phase, we use traditional SfM and MVS methods to reconstruct a virtual scene roughly matching the real scene. Then in the optimization phase, we adopt a hybrid approach to refine the geometry and reflectance, where the geometry is first optimized using an approximate differentiable rendering method, and the reflectance is optimized afterward using a physically-based differentiable rendering method. Our hybrid approach combines the efficiency of approximate methods with the high-quality results of physically-based methods. Extensive experiments on synthetic and real data demonstrate that our method can produce reconstructions with similar or higher quality than state-of-the-art methods while being more efficient.

The robustness of deep neural networks in safety-critical systems has received significant interest recently, which measures how sensitive the model output is under input perturbations. While most previous works focused on the local robustness property, the studies of the global robustness property, i.e., the robustness in the entire input space, are still lacking. In this work, we formulate the global robustness certification problem for ReLU neural networks and present an efficient approach to address it. Our approach includes a novel interleaving twin-network encoding scheme and an over-approximation algorithm leveraging relaxation and refinement techniques. Its timing efficiency and effectiveness are evaluated and compared with other state-of-the-art global robustness certification methods, and demonstrated via case studies on practical applications.

Minimizing a convex function with access to a first order oracle---that returns the function evaluation and (sub)gradient at a query point---is a canonical optimization problem and a fundamental primitive in machine learning. Gradient-based methods are the most popular approaches used for solving the problem, owing to their simplicity and computational efficiency. These methods, however, do not achieve the information-theoretically optimal query complexity for minimizing the underlying function to small error, which are achieved by more expensive techniques based on cutting-plane methods. Is it possible to achieve the information-theoretically query complexity without using these more complex and computationally expensive methods? In this work, we use memory as a lens to understand this, and show that is is not possible to achieve optimal query complexity without using significantly more memory than that used by gradient descent.

We present a method for extracting general modules for ontologies formulated in the description logic ALC. A module for an ontology is an ideally substantially smaller ontology that preserves all entailments for a user-specified set of terms. As such, it has applications such as ontology reuse and ontology analysis. Different from classical modules, general modules may use axioms not explicitly present in the input ontology, which allows for additional conciseness. So far, general modules have only been investigated for lightweight description logics. We present the first work that considers the more expressive description logic ALC. In particular, our contribution is a new method based on uniform interpolation supported by some new theoretical results. Our evaluation indicates that our general modules are often smaller than classical modules and uniform interpolants computed by the state-of-the-art, and compared with uniform interpolants, can be computed in significantly shorter time. Moreover, our method can be used for, and in fact, improves the computation of uniform interpolants and classical modules.

Recently, group link prediction has received increasing attention due to its important role in analyzing relationships between individuals and groups. However, most existing group link prediction methods emphasize static settings or only make cursory exploitation of historical information, so they fail to obtain good performance in dynamic applications. To this end, we attempt to solve the group link prediction problem in continuous-time dynamic scenes with fine-grained temporal information. We propose a novel continuous-time group link prediction method CTGLP to capture the patterns of future link formation between individuals and groups. A new graph neural network CTGNN is presented to learn the latent representations of individuals by biasedly aggregating neighborhood information. Moreover, we design an importance-based group modeling function to model the embedding of a group based on its known members. CTGLP eventually learns a probability distribution and predicts the link target. Experimental results on various datasets with and without unseen nodes show that CTGLP outperforms the state-of-the-art methods by 13.4% and 13.2% on average.

The scarcity of labeled data is a long-standing challenge for many machine learning tasks. We propose our gradient flow method to leverage the existing dataset (i.e., source) to generate new samples that are close to the dataset of interest (i.e., target). We lift both datasets to the space of probability distributions on the feature-Gaussian manifold, and then develop a gradient flow method that minimizes the maximum mean discrepancy loss. To perform the gradient flow of distributions on the curved feature-Gaussian space, we unravel the Riemannian structure of the space and compute explicitly the Riemannian gradient of the loss function induced by the optimal transport metric. For practical applications, we also propose a discretized flow, and provide conditional results guaranteeing the global convergence of the flow to the optimum. We illustrate the results of our proposed gradient flow method on several real-world datasets and show our method can improve the accuracy of classification models in transfer learning settings.

Decentralized multi-agent cooperative learning is a practical task due to the partially observed setting both in training and execution. Every agent learns to cooperate without access to the observations and policies of others. However, the decentralized training of multi-agent is of great difficulty due to non-stationarity, especially when other agents' policies are also in learning during training. To overcome this, we propose to learn a dynamic policy belief for each agent to predict the current policies of other agents and accordingly condition the policy of its own. To quickly adapt to the development of others' policies, we introduce a historical context to learn the belief inference according to a few recent action histories of other agents and a latent variational inference to model their policies by a learned distribution. We evaluate our method on the StarCraft II micro management task (SMAC) and demonstrate its superior performance in the decentralized training settings and comparable results with the state-of-the-art CTDE methods.

Mammogram image is important for breast cancer screening, and typically obtained in a dual-view form, i.e., cranio-caudal (CC) and mediolateral oblique (MLO), to provide complementary information for clinical decisions. However, previous methods mostly learn features from the two views independently, which violates the clinical knowledge and ignores the importance of dual-view correlation in the feature learning. In this paper, we propose a dual-view correlation hybrid attention network (DCHA-Net) for robust holistic mammogram classification. Specifically, DCHA-Net is carefully designed to extract and reinvent deep feature maps for the two views, and meanwhile to maximize the underlying correlations between them. A hybrid attention module, consisting of local relation and non-local attention blocks, is proposed to alleviate the spatial misalignment of the paired views in the correlation maximization. A dual-view correlation loss is introduced to maximize the feature similarity between corresponding strip-like regions with equal distance to the chest wall, motivated by the fact that their features represent the same breast tissues, and thus should be highly-correlated with each other. Experimental results on the two public datasets, i.e., INbreast and CBIS-DDSM, demonstrate that the DCHA-Net can well preserve and maximize feature correlations across views, and thus outperforms previous state-of-the-art methods for classifying a whole mammogram as malignant or not.

Video summarization and video captioning both condense the video content from the perspective of visual and text modes, i.e. the keyframe selection and language description generation. Existing video-and-language learning models commonly sample multiple frames for training instead of observing all. These sampled deputies greatly improve computational efficiency, but do they represent the original video content enough with no more redundancy? In this work, we propose a dual video summarization framework and verify it in the context of video captioning. Given the video frames, we firstly extract the visual representation based on the ViT model fine-tuned on the video-text domain. Then we summarize the keyframes according to the frame-lever score. To compress the number of keyframes as much as possible while ensuring the quality of captioning, we learn a cross-modal video summarizer to select the most semantically consistent frames according to the pseudo score label. Top K frames ( K is no more than 3% of the entire video.) are chosen to form the video representation. Moreover, to evaluate the static appearance and temporal information of video, we design the ranking scheme of video representation from two aspects: feature-oriented and sequence-oriented. Finally, we generate the descriptions with a lightweight LSTM decoder. The experiment results on the MSR-VTT and MSVD dataset reveal that, for the generative task as video captioning, a small number of keyframes can convey the same semantic information to perform well on captioning, or even better than the original sampling.

Knowledge distillation is an effective method for model compression. However, it is still a challenging topic to apply knowledge distillation to detection tasks. There are two key points resulting in poor distillation performance for detection tasks. One is the serious imbalance between foreground and background features, another one is that small object lacks enough feature representation. To solve the above issues, we propose a new distillation method named dual relation knowledge distillation (DRKD), including pixel-wise relation distillation and instance-wise relation distillation. The pixel-wise relation distillation embeds pixel-wise features in the graph space and applies graph convolution to capture the global pixel relation. By distilling the global pixel relation, the student detector can learn the relation between foreground and background features, and avoid the difficulty of distilling features directly for the feature imbalance issue. Besides, we find that instance-wise relation supplements valuable knowledge beyond independent features for small objects. Thus, the instance-wise relation distillation is designed, which calculates the similarity of different instances to obtain a relation matrix. More importantly, a relation filter module is designed to highlight valuable instance relations. The proposed dual relation knowledge distillation is general and can be easily applied for both one-stage and two-stage detectors. Our method achieves state-of-the-art performance, which improves Faster R-CNN based on ResNet50 from 38.4% to 41.6% mAP and improves RetinaNet based on ResNet50 from 37.4% to 40.3% mAP on COCO 2017.

Recent efforts have achieved remarkable progress on single image deraining on the stationary distributed data. However, catastrophic forgetting raises practical concerns when applying these methods to real applications, where the data distributions change constantly. In this paper, we investigate the continual learning issue for rain removal and develop a novel efficient continual learned deraining transformer. Different from the typical replay or regularization-based methods that increase overall training time or parameter space, our method relies on compact prompts which are learnable parameters, to maintain both task-invariant and task-specific knowledge. Our prompts are applied at both image and feature levels to leverage effectively transferred knowledge of images and features among different tasks. We conduct comprehensive experiments under widely-used rain removal datasets, where our proposed dual prompt learning consistently outperforms prior state-of-the-art methods. Moreover, we observe that, even though our method is designed for continual learning, it still achieves superior results on the stationary distributed data, which further demonstrates the effectiveness of our method. Our website is available at: http://liuminghao.com.cn/DPL/.

Federated recommendation is a new Internet service architecture that aims to provide privacy-preserving recommendation services in federated settings. Existing solutions are used to combine distributed recommendation algorithms and privacy-preserving mechanisms. Thus it inherently takes the form of heavyweight models at the server and hinders the deployment of on-device intelligent models to end-users. This paper proposes a novel Personalized Federated Recommendation (PFedRec) framework to learn many user-specific lightweight models to be deployed on smart devices rather than a heavyweight model on a server. Moreover, we propose a new dual personalization mechanism to effectively learn fine-grained personalization on both users and items. The overall learning process is formulated into a unified federated optimization framework. Specifically, unlike previous methods that share exactly the same item embeddings across users in a federated system, dual personalization allows mild finetuning of item embeddings for each user to generate user-specific views for item representations which can be integrated into existing federated recommendation methods to gain improvements immediately. Experiments on multiple benchmark datasets have demonstrated the effectiveness of PFedRec and the dual personalization mechanism. Moreover, we provide visualizations and in-depth analysis of the personalization techniques in item embedding, which shed novel insights on the design of recommender systems in federated settings. The code is available.

Organic reaction prediction is a critical task in drug discovery. Recently, researchers have achieved non-autoregressive reaction prediction by modeling the redistribution of electrons, resulting in state-of-the-art top-1 accuracy, and enabling parallel sampling. However, the current non-autoregressive decoder does not satisfy two essential rules of electron redistribution modeling simultaneously: the electron-counting rule and the symmetry rule. This violation of the physical constraints of chemical reactions impairs model performance. In this work, we propose a new framework called ReactionSink that combines two doubly stochastic self-attention mappings to obtain electron redistribution predictions that follow both constraints. We further extend our solution to a general multi-head attention mechanism with augmented constraints. To achieve this, we apply Sinkhorn's algorithm to iteratively update self-attention mappings, which imposes doubly conservative constraints as additional informative priors on electron redistribution modeling. We theoretically demonstrate that our ReactionSink can simultaneously satisfy both rules, which the current decoder mechanism cannot do. Empirical results show that our approach consistently improves the predictive performance of non-autoregressive models and does not bring an unbearable additional computational cost.