IJCAI 2023 Papers

Neural Architecture Search (NAS) aims to automatically excavate the optimal network architecture with superior test performance. Recent neural architecture search (NAS) approaches rely on validation loss or accuracy to find the superior network for the target data. In this paper, we investigate a new neural architecture search measure for excavating architectures with better generalization. We demonstrate that the flatness of the loss surface can be a promising proxy for predicting the generalization capability of neural network architectures. We evaluate our proposed method on various search spaces, showing similar or even better performance compared to the state-of-the-art NAS methods. Notably, the resultant architecture found by flatness measure generalizes robustly to various shifts in data distribution (e.g. ImageNet-V2,-A,-O), as well as various tasks such as object detection and semantic segmentation.

Graph Transformers (GTs) have proved their advantage in graph-level tasks. However, existing GTs still perform unsatisfactorily on the node classification task due to 1) the overwhelming unrelated information obtained from a vast number of irrelevant distant nodes and 2) the quadratic complexity regarding the number of nodes via the fully connected attention mechanism. In this paper, we present Gapformer, a method for node classification that deeply incorporates Graph Transformer with Graph Pooling. More specifically, Gapformer coarsens the large-scale nodes of a graph into a smaller number of pooling nodes via local or global graph pooling methods, and then computes the attention solely with the pooling nodes rather than all other nodes. In such a manner, the negative influence of the overwhelming unrelated nodes is mitigated while maintaining the long-range information, and the quadratic complexity is reduced to linear complexity with respect to the fixed number of pooling nodes. Extensive experiments on 13 node classification datasets, including homophilic and heterophilic graph datasets, demonstrate the competitive performance of Gapformer over existing Graph Neural Networks and GTs.

Artificial Intelligence often relies on information obtained from others through crowdsourcing, federated learning, or data markets. It is crucial to ensure that this data is accurate. Over the past 20 years, a variety of incentive mechanisms have been developed that use game theory to reward the accuracy of contributed data. These techniques are applicable to many settings where AI uses contributed data. This survey categorizes the different techniques and their properties, and shows their limits and tradeoffs. It identifies open issues and points to possible directions to address these.

Game-theoretic interactions with AI agents could differ from traditional human-human interactions in various ways. One such difference is that it may be possible to simulate an AI agent (for example because its source code is known), which allows others to accurately predict the agent's actions. This could lower the bar for trust and cooperation. In this paper, we first formally define games in which one player can simulate another at a cost, and derive some basic properties of such games. Then, we prove a number of results for such games, including: (1) introducing simulation into generic-payoff normal-form games makes them easier to solve; (2) if the only obstacle to cooperation is a lack of trust in the possibly-simulated agent, simulation enables equilibria that improve the outcome for both agents; and (3) however, there are settings where introducing simulation results in strictly worse outcomes for both players.

Dynamic scene graph generation aims to identify visual relationships (subject-predicate-object) in frames based on spatio-temporal contextual information in the video. Previous work implicitly models the spatio-temporal interaction simultaneously, which leads to entanglement of spatio-temporal contextual information. To this end, we propose a Grafting-Then-Reassembling framework (GTR), which explicitly extracts intra-frame spatial information and inter-frame temporal information in two separate stages to decouple spatio-temporal contextual information. Specifically, we first graft a static scene graph generation model to generate static visual relationships within frames. Then we propose the temporal dependency model to extract the temporal dependencies across frames, and explicitly reassemble static visual relationships into dynamic scene graphs. Experimental results show that GTR achieves the state-of-the-art performance on Action Genome dataset. Further analyses reveal that the reassembling stage is crucial to the success of our framework.

Optimizing molecules with desired properties is a crucial step in de novo drug design. While translation-based methods have achieved initial success, they continue to face the challenge of the “exposure bias” problem. The challenge of preventing the “exposure bias” problem of molecule optimization lies in the need for both positive and negative molecules of contrastive learning. That is because generating positive molecules through data augmentation requires domain-specific knowledge, and randomly sampled negative molecules are easily distinguished from the real molecules. Hence, in this work, we propose a molecule optimization method called GPMO, which leverages a gradient perturbation-based contrastive learning method to prevent the “exposure bias” problem in translation-based molecule optimization. With the assistance of positive and negative molecules, GPMO is able to effectively handle both real and artificial molecules. GPMO is a molecule optimization method that is conditioned on matched molecule pairs for drug discovery. Our empirical studies show that GPMO outperforms the state-of-the- art molecule optimization methods. Furthermore, the negative and positive perturbations improve the robustness of GPMO.

The use of multi-agent reinforcement learning (MARL) methods in coordinating traffic lights (CTL) has become increasingly popular, treating each intersection as an agent. However, existing MARL approaches either treat each agent absolutely homogeneous, i.e., same network and parameter for each agent, or treat each agent completely heterogeneous, i.e., different networks and parameters for each agent. This creates a difficult balance between accuracy and complexity, especially in large-scale CTL. To address this challenge, we propose a grouped MARL method named GPLight. We first mine the similarity between agent environment considering both real-time traffic flow and static fine-grained road topology. Then we propose two loss functions to maintain a learnable and dynamic clustering, one that uses mutual information estimation for better stability, and the other that maximizes separability between groups. Finally, GPLight enforces the agents in a group to share the same network and parameters. This approach reduces complexity by promoting cooperation within the same group of agents while reflecting differences between groups to ensure accuracy. To verify the effectiveness of our method, we conduct experiments on both synthetic and real-world datasets, with up to 1,089 intersections. Compared with state-of-the-art methods, experiment results demonstrate the superiority of our proposed method, especially in large-scale CTL.

Protein binding site prediction is an important prerequisite for the discovery of new drugs. Usually, natural 3D U-Net is adopted as the standard site prediction framework to do per-voxel binary mask classification. However, this scheme only performs feature extraction for single-scale samples, which may bring the loss of global or local information, resulting in incomplete, artifacted or even missed predictions. To tackle this issue, we propose a network called GLPocket, which is based on the Lmser (Least mean square error reconstruction) network and utilizes multi-scale representation to predict binding sites. Firstly, GLPocket uses Target Cropping Block (TCB) for targeted prediction. TCB selects the local interested feature from the global representations to perform concentrated prediction, and reduces the volume of feature maps to be calculated by 82% without adding additional parameters. It integrates global distribution information into local regions, making prediction more concentrated on decoding stage. Secondly, GLPocket establishes long-range relationship of patches within the local region with Transformer Block (TB), to enrich local context semantic information. Experiments show that GLPocket improves by 0.5%-4% on DCA Top-n prediction compared with previous state-of-the-art methods on four datasets. Our code has been released in https://github.com/CMACH508/GLPocket.

In a number of different fields, including Engeneering, Chemistry and Physics, the design of technological tools and device structures is increasingly supported by deep-learning based methods, which provide suggestions on crucial architectural choices based on the properties that these tools and structures should exhibit. The paper proposes a novel architecture, named GIDnet, to address this inverse design problem, which is based on exploring a suitably defined latent space associated with the possible designs. Among its distinguishing features, GIDnet is capable of identifying the most appropriate starting point for the exploration and of likely converging into a point corresponding to a design that is a feasible one. Results of a thorough experimental activity evidence that GIDnet outperforms earlier approaches in the literature.

Node classification is the task of predicting the labels of unlabeled nodes in a graph. State-of-the-art methods based on graph neural networks achieve excellent performance when all labels are available during training. But in real-life, models are of ten applied on data with new classes, which can lead to massive misclassification and thus significantly degrade performance. Hence, developing open-set classification methods is crucial to determine if a given sample belongs to a known class. Existing methods for open-set node classification generally use transductive learning with part or all of the features of real unseen class nodes to help with open-set classification. In this paper, we propose a novel generative open-set node classification method, i.e., G2Pxy, which follows a stricter inductive learning setting where no information about unknown classes is available during training and validation. Two kinds of proxy unknown nodes, inter-class unknown proxies and external unknown proxies are generated via mixup to efficiently anticipate the distribution of novel classes. Using the generated proxies, a closed-set classifier can be transformed into an open-set one, by augmenting it with an extra proxy classifier. Under the constraints of both cross entropy loss and complement entropy loss, G2Pxy achieves superior effectiveness for unknown class detection and known class classification, which is validated by experiments on bench mark graph datasets. Moreover, G2Pxy does not have specific requirement on the GNN architecture and shows good generalizations.

The industrial multi-generator Wave Energy Converters (WEC) must handle multiple simultaneous waves coming from different directions called spread waves. These complex devices in challenging circumstances need controllers with multiple objectives of energy capture efficiency, reduction of structural stress to limit maintenance, and proactive protection against high waves. The Multi-Agent Reinforcement Learning (MARL) controller trained with Proximal Policy Optimization (PPO) algorithm can handle these complexities. In this paper, we explore different function approximations for the policy and critic networks in modeling the sequential nature of the system dynamics and find that they are key to better performance. We investigated the performance of a fully connected neural network (FCN), LSTM, and Transformer model variants with varying depths and gated residual connections. Our results show that the transformer model of moderate depth with gated residual connections around the multi-head attention, multi-layer perceptron, and the transformer block (STrXL) proposed in this paper is optimal and boosts energy efficiency by an average of 22.1% for these complex spread waves over the existing spring damper (SD) controller. Furthermore, unlike the default SD controller, the transformer controller almost eliminated the mechanical stress from the rotational yaw motion for angled waves. Demo: https://tinyurl.com/yueda3jh

The rapid rise in cloud computing has resulted in an alarming increase in data centers' carbon emissions, which now accounts for >3% of global greenhouse gas emissions, necessitating immediate steps to combat their mounting strain on the global climate. An important focus of this effort is to improve resource utilization in order to save electricity usage. Our proposed Full Scaling Automation (FSA) mechanism is an effective method of dynamically adapting resources to accommodate changing workloads in large-scale cloud computing clusters, enabling the clusters in data centers to maintain their desired CPU utilization target and thus improve energy efficiency. FSA harnesses the power of deep representation learning to accurately predict the future workload of each service and automatically stabilize the corresponding target CPU usage level, unlike the previous autoscaling methods, such as Autopilot or FIRM, that need to adjust computing resources with statistical models and expert knowledge. Our approach achieves significant performance improvement compared to the existing work in real-world datasets. We also deployed FSA on large-scale cloud computing clusters in industrial data centers, and according to the certification of the China Environmental United Certification Center (CEC), a reduction of 947 tons of carbon dioxide, equivalent to a saving of 1538,000 kWh of electricity, was achieved during the Double 11 shopping festival of 2022, marking a critical step for our company’s strategic goal towards carbon neutrality by 2030.

Recent research on bidirectional heuristic search (BiHS) is based on the must-expand pairs theory (MEP theory), which describes which pairs of nodes must be expanded during the search to guarantee the optimality of solutions. A separate line of research in BiHS has proposed algorithms that use lower bounds that are derived from consistent heuristics during search. This paper links these two directions, providing a comprehensive unifying view and showing that both existing and novel algorithms can be derived from the MEP theory. An extended set of bounds is formulated, encompassing both previously discovered bounds and new ones. Finally, the bounds are empirically evaluated by their contribution to the efficiency of the search

Most existing Low-Light Image Enhancement (LLIE) methods are primarily designed to improve brightness in dark regions, which suffer from severe degradation in nighttime images. However, these methods have limited exploration in another major visibility damage, the glow effects in real night scenes. Glow effects are inevitable in the presence of artificial light sources and cause further diffused blurring when directly enhanced. To settle this issue, we innovatively consider the glow suppression task as learning physical glow generation via multiple scattering estimation according to the Atmospheric Point Spread Function (APSF). In response to the challenges posed by uneven glow intensity and varying source shapes, an APSF-based Nighttime Imaging Model with Near-field Light Sources (NIM-NLS) is specifically derived to design a scalable Light-aware Blind Deconvolution Network (LBDN). The glow-suppressed result is then brightened via a Retinex-based Enhancement Module (REM). Remarkably, the proposed glow suppression method is based on zero-shot learning and does not rely on any paired or unpaired training data. Empirical evaluations demonstrate the effectiveness of the proposed method in both glow suppression and low-light enhancement tasks.

With the development of Vision-Language Pre-training Models (VLPMs) represented by CLIP and ALIGN, significant breakthroughs have been achieved for association-based visual tasks such as image classification and image-text retrieval by the zero-shot capability of CLIP without fine-tuning. However, CLIP is hard to apply to generation-based tasks. This is due to the lack of decoder architecture and pre-training tasks for generation. Although previous works have created generation capacity for CLIP through additional language models, a modality gap between the CLIP representations of different modalities and the inability of CLIP to model the offset of this gap, which results in the failure of the concept to transfer across modes. To solve the problem, we try to map images/videos to the language modality and generate captions from the language modality. In this paper, we propose the K-nearest-neighbor Cross-modality Mapping (Knight), a zero-shot method from association to generation. With vision-free unsupervised training, Knight achieves state-of-the-art performance in zero-shot methods for image captioning and video captioning.

Deep learning is increasingly used to learn policies for planning problems, yet policies represented by neural networks are difficult to interpret, verify and trust. Existing formal approaches to post-hoc explanations provide concise reasons for a single decision made by an ML model. However, understanding planning policies require explaining sequences of decisions. In this paper, we formulate the problem of finding explanations for the sequence of decisions recommended by a learnt policy in a given state. We show that, under certain assumptions, a minimal explanation for a sequence can be computed by solving a number of single decision explanation problems which is linear in the length of the sequence. We present experimental results of our implementation of this approach for ASNet policies for classical planning domains.

Agriculture faces unprecedented challenges due to climate change, population growth, and water scarcity. These challenges highlight the need for efficient resource usage to optimize crop production. Conventional techniques for forecasting hydrological response features, such as soil moisture, rely on physics-based and empirical hydrological models, which necessitate significant time and domain expertise. Drawing inspiration from traditional hydrological modeling, a novel temporal graph convolution neural network has been constructed. This involves grouping units based on their time-varying hydrological properties, constructing graph topologies for each cluster based on similarity using dynamic time warping, and utilizing graph convolutions and a gated recurrent neural network to forecast soil moisture. The method has been trained, validated, and tested on field-scale time series data spanning 40 years in northeastern United States. Results show that using domain-inspired clustering with time series graph neural networks is more effective in forecasting soil moisture than existing models. This framework is being deployed as part of a pro bono social impact program that leverages hybrid cloud and AI technologies to enhance and scale non-profit and government organizations. The trained models are currently being deployed on a series of small-holding farms in central Texas.

In this paper, we present a computational analysis of the Persian language Twitter discourse with the aim to estimate the shift in stance toward gender equality following the death of Mahsa Amini in police custody. We present an ensemble active learning pipeline to train a stance classifier. Our novelty lies in the involvement of Iranian women in an active role as annotators in building this AI system. Our annotators not only provide labels, but they also suggest valuable keywords for more meaningful corpus creation as well as provide short example documents for a guided sampling step. Our analyses indicate that Mahsa Amini's death triggered polarized Persian language discourse where both fractions of negative and positive tweets toward gender equality increased. The increase in positive tweets was slightly greater than the increase in negative tweets. We also observe that with respect to account creation time, between the state-aligned Twitter accounts and pro-protest Twitter accounts, pro-protest accounts are more similar to baseline Persian Twitter activity.

Most infrared small target detection (ISTD) networks focus on building effective neural blocks or feature fusion modules but none describes the ISTD process from the image evolution perspective. The directional evolution of image pixels influenced by convolution, pooling and surrounding pixels is analogous to the movement of fluid elements constrained by surrounding variables ang particles. Inspired by this, we explore a novel research routine by abstracting the movement of pixels in the ISTD process as the flow of fluid in fluid dynamics (FD). Specifically, a new Fluid Dynamics-Inspired Network (FDI-Net) is devised for ISTD. Based on Taylor Central Difference (TCD) method, the TCD feature extraction block is designed, where convolution and Transformer structures are combined for local and global information. The pixel motion equation during the ISTD process is derived from the Navier–Stokes (N-S) equation, constructing a N-S Refinement Module that refines extracted features with edge details. Thus, the TCD feature extraction block determines the primary movement direction of pixels during detection, while the N-S Refinement Module corrects some skewed directions of the pixel stream to supplement the edge details. Experiments on IRSTD-1k and SIRST demonstrate that our method achieves SOTA performance in terms of evaluation metrics.