IJCAI 2022 论文列表

In this work we consider the problem of how to best allocate a limited supply of vaccines in the aftermath of an infectious disease outbreak by viewing the problem as a sequential game between a learner and an environment (specifically, a bandit problem). The difficulty of this problem lies in the fact that the payoff of vaccination cannot be directly observed, making it difficult to compare the relative effectiveness of vaccination on different population groups. Currently used vaccination policies make recommendations based on mathematical modelling and ethical considerations. These policies are static, and do not adapt as conditions change. Our aim is to design and evaluate an algorithm which can make use of routine surveillance data to dynamically adjust its recommendation. We evaluate the performance of our approach by applying it to a simulated epidemic of a disease based on real-world COVID-19 data, and show that our vaccination policy was able to perform better than existing vaccine allocation policies. In particular, we show that with our allocation method, we can reduce the number of required vaccination by at least 50% in order to keep the peak number of hospitalised patients below a certain threshold. Also, when the same batch sizes are used, our method can reduce the peak number of hospitalisation by up to 20%. We also demonstrate that our vaccine allocation does not vary the number of batches per group much, making it socially more acceptable (as it reduces uncertainty, hence results in better and more interpretable communication).

We present Bayesian Team Imitation Learner (BTIL), an imitation learning algorithm to model the behavior of teams performing sequential tasks in Markovian domains. In contrast to existing multi-agent imitation learning techniques, BTIL explicitly models and infers the time-varying mental states of team members, thereby enabling learning of decentralized team policies from demonstrations of suboptimal teamwork. Further, to allow for sample- and label-efficient policy learning from small datasets, BTIL employs a Bayesian perspective and is capable of learning from semi-supervised demonstrations. We demonstrate and benchmark the performance of BTIL on synthetic multi-agent tasks as well as a novel dataset of human-agent teamwork. Our experiments show that BTIL can successfully learn team policies from demonstrations despite the influence of team members' (time-varying and potentially misaligned) mental states on their behavior.

Generative methods have been successfully applied in zero-shot learning (ZSL) by learning an implicit mapping to alleviate the visual-semantic domain gaps and synthesizing unseen samples to handle the data imbalance between seen and unseen classes. However, existing generative methods simply use visual features extracted by the pre-trained CNN backbone. These visual features lack attribute-level semantic information. Consequently, seen classes are indistinguishable, and the knowledge transfer from seen to unseen classes is limited. To tackle this issue, we propose a novel Semantic Compression Embedding Guided Generation (SC-EGG) model, which cascades a semantic compression embedding network (SCEN) and an embedding guided generative network (EGGN). The SCEN extracts a group of attribute-level local features for each sample and further compresses them into the new low-dimension visual feature. Thus, a dense-semantic visual space is obtained. The EGGN learns a mapping from the class-level semantic space to the dense-semantic visual space, thus improving the discriminability of the synthesized dense-semantic unseen visual features. Extensive experiments on three benchmark datasets, i.e., CUB, SUN and AWA2, demonstrate the significant performance gains of SC-EGG over current state-of-the-art methods and its baselines.

Unsupervised semantic segmentation requires assigning a label to every pixel without any human annotations. Despite recent advances in self-supervised representation learning for individual images, unsupervised semantic segmentation with pixel-level representations is still a challenging task and remains underexplored. In this work, we propose a self-supervised pixel representation learning method for semantic segmentation by using visual concepts (i.e., groups of pixels with semantic meanings, such as parts, objects, and scenes) extracted from images. To guide self-supervised learning, we leverage three types of relationships between pixels and concepts, including the relationships between pixels and local concepts, local and global concepts, as well as the co-occurrence of concepts. We evaluate the learned pixel embeddings and visual concepts on three datasets, including PASCAL VOC 2012, COCO 2017, and DAVIS 2017. Our results show that the proposed method gains consistent and substantial improvements over recent unsupervised semantic segmentation approaches, and also demonstrate that visual concepts can reveal insights into image datasets.

Existing deep image dehazing methods usually depend on supervised learning with a large number of hazy-clean image pairs which are expensive or difficult to collect. Moreover, dehazing performance of the learned model may deteriorate significantly when the training hazy-clean image pairs are insufficient and are different from real hazy images in applications. In this paper, we show that exploiting large scale training set and adapting to real hazy images are two critical issues in learning effective deep dehazing models. Under the depth guidance estimated by a well-trained depth estimation network, we leverage the conventional atmospheric scattering model to generate massive hazy-clean image pairs for the self-supervised pre-training of dehazing network. Furthermore, self-supervised adaptation is presented to adapt pre-trained network to real hazy images. Learning without forgetting strategy is also deployed in self-supervised adaptation by combining self-supervision and model adaptation via contrastive learning. Experiments show that our proposed method performs favorably against the state-of-the-art methods, and is quite efficient, i.e., handling a 4K image in 23 ms. The codes are available at https://github.com/DongLiangSXU/SLAdehazing.

Traditional recommendation usually focuses on utilizing only one target user behavior (e.g., purchase) but ignoring other auxiliary behaviors (e.g., click, add to cart). Early efforts of multi-behavior recommendation often emphasize the differences between multiple behaviors, i.e., they aim to extract useful information by distinguishing different behaviors. However, the commonality between them, which reflects user's common preference for items associated with different behaviors, is largely ignored. Meanwhile, the multi-behavior recommendation still severely suffers from limited supervision signal issue. In this paper, we propose a novel self-supervised graph collaborative filtering model for multi-behavior recommendation named S-MBRec. Specifically, for each behavior, we execute the GCNs to learn the user and item embeddings. Then we design a supervised task, distinguishing the importance of different behaviors, to capture the differences between embeddings. Meanwhile, we propose a star-style contrastive learning task to capture the embedding commonality between target and auxiliary behaviors, so as to alleviate the sparsity of supervision signal, reduce the redundancy among auxiliary behavior, and extract the most critical information. Finally, we jointly optimize the above two tasks. Extensive experiments, in comparison with state-of-the-arts, well demonstrate the effectiveness of S-MBRec, where the maximum improvement can reach to 20%.

Multi-source domain adaptation has attracted great attention in machine learning community. Most of these methods focus on weighting the predictions produced by the adaptation networks of different domains. Thus the domain shifts between certain of domains and target domain are not effectively relieved, resulting in that these domains are not fully exploited and even may have a negative influence on multi-source domain adaptation task. To address such challenge, we propose a multi-source domain adaptation method to gradually improve the adaptation ability of each source domain by producing more high-confident pseudo-labels with self-paced learning for conditional distribution alignment. The proposed method first trains several separate domain branch networks with single domains and an ensemble branch network with all domains. Then we obtain some high-confident pseudo-labels with the branch networks and learn the branch specific pseudo-labels with self-paced learning. Each branch network reduces the domain gap by aligning the conditional distribution with its branch specific pseudo-labels and the pseudo-labels provided by all branch networks. Experiments on Office31, Office-Home and DomainNet show that the proposed method outperforms the state-of-the-art methods.

Mimicking the sampling mechanism of the primate fovea, a retina-inspired vision sensor named spiking camera has been developed, which has shown great potential for capturing high-speed dynamic scenes with a sampling rate of 40,000 Hz. Unlike conventional digital cameras, the spiking camera continuously captures photons and outputs asynchronous binary spikes with various inter-spike intervals to record dynamic scenes. However, how to reconstruct dynamic scenes from asynchronous spike streams remains challenging. In this work, we propose a novel pretext task to build a self-supervised reconstruction framework for spiking cameras. Specifically, we utilize the blind-spot network commonly used in self-supervised denoising tasks as our backbone, and perform self-supervised learning by constructing proper pseudo-labels. In addition, in view of the poor scalability and insufficient information utilization of the blind-spot network, we present a mutual learning framework to improve the overall performance of the network through mutual distillation between a non-blind-spot network and a blind-spot network. This also enables the network to bypass constraints of the blind-spot network, allowing state-of-the-art modules to be used to further improve performance. The experimental results demonstrate that our methods evidently outperform previous unsupervised spiking camera reconstruction methods and achieve desirable results compared with supervised methods.

Sleep staging is an important task that enables sleep quality assessment and disorder diagnosis. Due to dependency on manually labeled data, many researches have turned from supervised approaches to self-supervised learning (SSL) for sleep staging. While existing SSL methods have made significant progress in terms of its comparable performance to supervised methods, there are still some limitations. Contrastive learning could potentially lead to false negative pair assignments in sleep signal data. Moreover, existing data augmentation techniques directly modify the original signal data, making it likely to lose important information. To mitigate these issues, we propose Self-Supervised Learning with Attention-aided Positive Pairs (SSLAPP). Instead of the contrastive learning, SSLAPP carefully draws high-quality positive pairs and exploits them in representation learning. Here, we propose attention-based latent signal augmentation, which plays a key role by capturing important features without losing valuable signal information. Experimental results show that our proposed method achieves state-of-the-art performance in sleep stage classification with limited labeled data. The code is available at: https://github.com/DILAB-HYU/SSLAPP

Vision-based reinforcement learning requires efficient and robust representations of image-based observations, especially when the images contain distracting (task-irrelevant) elements such as shadows, clouds, and light. It becomes more important if those distractions are not exposed during training. We design a Self-Predictive Dynamics (SPD) method to extract task-relevant features efficiently, even in unseen observations after training. SPD uses weak and strong augmentations in parallel, and learns representations by predicting inverse and forward transitions across the two-way augmented versions. In a set of MuJoCo visual control tasks and an autonomous driving task (CARLA), SPD outperforms previous studies in complex observations, and significantly improves the generalization performance for unseen observations. Our code is available at https://github.com/unigary/SPD.

Recent unsupervised person re-identification (reID) methods mostly apply pseudo labels from clustering algorithms as supervision signals. Despite great success, this fashion is very likely to aggregate different identities with similar appearances into the same cluster. In result, the hard negative samples, playing important role in training reID models, are significantly reduced. To alleviate this problem, we propose a self-guided hard negative generation method for unsupervised person re-ID. Specifically, a joint framework is developed which incorporates a hard negative generation network (HNGN) and a re-ID network. To continuously generate harder negative samples to provide effective supervisions in the contrastive learning, the two networks are alternately trained in an adversarial manner to improve each other, where the reID network guides HNGN to generate challenging data and HNGN enforces the re-ID network to enhance discrimination ability. During inference, the performance of re-ID network is improved without introducing any extra parameters. Extensive experiments demonstrate that the proposed method significantly outperforms a strong baseline and also achieves better results than state-of-the-art methods.

Input distribution shift is one of the vital problems in unsupervised domain adaptation (UDA). The most popular UDA approaches focus on domain-invariant representation learning, trying to align the features from different domains into a similar feature distribution. However, these approaches ignore the direct alignment of input word distributions between domains, which is a vital factor in word-level classification tasks such as cross-domain NER. In this work, we shed new light on cross-domain NER by introducing a subword-level solution, X-Piece, for input word-level distribution shift in NER. Specifically, we re-tokenize the input words of the source domain to approach the target subword distribution, which is formulated and solved as an optimal transport problem. As this approach focuses on the input level, it can also be combined with previous DIRL methods for further improvement. Experimental results show the effectiveness of the proposed method based on BERT-tagger on four benchmark NER datasets. Also, the proposed method is proved to benefit DIRL methods such as DANN.

The development of AlphaZero was a breakthrough in search-based reinforcement learning, by employing a given world model in a Monte-Carlo tree search (MCTS) algorithm to incrementally learn both an action policy and a value estimation. When extending this paradigm to the setting of simultaneous move games we find that the selection strategy of AlphaZero has theoretical shortcomings, including that convergence to a Nash equilibrium is not guaranteed. By analyzing these shortcomings, we find that the selection strategy corresponds to an approximated version of bandit linear optimization using Tsallis entropy regularization with α parameter set to zero, which is equivalent to log-barrier regularization. This observation allows us to refine the search method used by AlphaZero to obtain an algorithm that has theoretically optimal regret as well as superior empirical performance on our evaluation benchmark.

Evaluation of deep reinforcement learning (RL) is inherently challenging. Especially the opaqueness of learned policies and the stochastic nature of both agents and environments make testing the behavior of deep RL agents difficult. We present a search-based testing framework that enables a wide range of novel analysis capabilities for evaluating the safety and performance of deep RL agents. For safety testing, our framework utilizes a search algorithm that searches for a reference trace that solves the RL task. The backtracking states of the search, called boundary states, pose safety-critical situations. We create safety test-suites that evaluate how well the RL agent escapes safety-critical situations near these boundary states. For robust performance testing, we create a diverse set of traces via fuzz testing. These fuzz traces are used to bring the agent into a wide variety of potentially unknown states from which the average performance of the agent is compared to the average performance of the fuzz traces. We apply our search-based testing approach on RL for Nintendo's Super Mario Bros.

In the past decade, several systems for learning Answer Set Programs (ASP) have been proposed, including the recent FastLAS system. Compared to other state-of-the-art approaches to learning ASP, FastLAS is more scalable, as rather than computing the hypothesis space in full, it computes a much smaller subset relative to a given set of examples that is nonetheless guaranteed to contain an optimal solution to the task (called an OPT-sufficient subset). On the other hand, like many other Inductive Logic Programming (ILP) systems, FastLAS is designed to be run on a fixed learning task meaning that if new examples are discovered after learning, the whole process must be run again. In many real applications, data arrives in a stream. Rerunning an ILP system from scratch each time new examples arrive is inefficient. In this paper we address this problem by presenting IncrementalLAS, a system that uses a new technique, called hypothesis space expansion, to enable a FastLAS-like OPT-sufficient subset to be expanded each time new examples are discovered. We prove that this preserves FastLAS's guarantee of finding an optimal solution to the full task (including the new examples), while removing the need to repeat previous computations. Through our evaluation, we demonstrate that running IncrementalLAS on tasks updated with sequences of new examples is significantly faster than re-running FastLAS from scratch on each updated task.

Using machine-learned predictions to create algorithms with better approximation guarantees is a very fresh and active field. In this work, we study classic scheduling problems under the learning augmented setting. More specifically, we consider the problem of scheduling jobs with arbitrary release dates on a single machine and the problem of scheduling jobs with a common release date on multiple machines. Our objective is to minimize the sum of completion times. For both problems, we propose algorithms which use predictions for taking their decisions. Our algorithms are consistent -- i.e. when the predictions are accurate, the performances of our algorithms are close to those of an optimal offline algorithm--, and robust -- i.e. when the predictions are wrong, the performance of our algorithms are close to those of an online algorithm without predictions. In addition, we confirm the above theoretical bounds by conducting experimental evaluation comparing the proposed algorithms to the offline optimal ones for both the single and multiple machines settings.

Recently, a variety of vision transformers have been developed as their capability of modeling long-range dependency. In current transformer-based backbones for medical image segmentation, convolutional layers were replaced with pure transformers, or transformers were added to the deepest encoder to learn global context. However, there are mainly two challenges in a scale-wise perspective: (1) intra-scale problem: the existing methods lacked in extracting local-global cues in each scale, which may impact the signal propagation of small objects; (2) inter-scale problem: the existing methods failed to explore distinctive information from multiple scales, which may hinder the representation learning from objects with widely variable size, shape and location. To address these limitations, we propose a novel backbone, namely ScaleFormer, with two appealing designs: (1) A scale-wise intra-scale transformer is designed to couple the CNN-based local features with the transformer-based global cues in each scale, where the row-wise and column-wise global dependencies can be extracted by a lightweight Dual-Axis MSA. (2) A simple and effective spatial-aware inter-scale transformer is designed to interact among consensual regions in multiple scales, which can highlight the cross-scale dependency and resolve the complex scale variations. Experimental results on different benchmarks demonstrate that our Scale-Former outperforms the current state-of-the-art methods. The code is publicly available at: https://github.com/ZJUGiveLab/ScaleFormer.

We study the problem of designing scalable algorithms to find effective intervention strategies for controlling stochastic epidemic processes on networks. This is a common problem arising in agent based models for epidemic spread. Previous approaches to this problem focus on either heuristics with no guarantees or approximation algorithms that scale only to networks corresponding to county-sized populations, typically, with less than a million nodes. In particular, the mathematical-programming based approaches need to solve the Linear Program (LP) relaxation of the problem using an LP solver, which restricts the scalability of this approach. In this work, we overcome this restriction by designing an algorithm that adapts the multiplicative weights update (MWU) framework, along with the sample average approximation (SAA) technique, to approximately solve the linear program (LP) relaxation for the problem. To scale this approach further, we provide a memory-efficient algorithm that enables scaling to large networks, corresponding to country-size populations, with over 300 million nodes and 30 billion edges. Furthermore, we show that this approach provides near-optimal solutions to the LP in practice.

The goal of this work is to develop novel methods to solve the semiconductor fab scheduling problem. The problem can be modeled as a flexible job-shop with large instances and specific constraints related to special machine and job characteristics. To investigate the problem, we develop a tool to simulate small to large-scale instances of the problem. Using the simulator, we aim to develop new dispatching strategies using genetic programming and reinforcement learning.