AAAI 2025 Papers

Short text classification, as a research subtopic in natural language processing, is more challenging due to its semantic sparsity and insufficient labeled samples in practical scenarios. We propose a novel model named MI-DELIGHT for short text classification in this work. Specifically, it first performs multi-source information (i.e., statistical information, linguistic information, and factual information) exploration to alleviate the sparsity issues. Then, the graph learning approach is adopted to learn the representation of short texts, which are presented in graph forms. Moreover, we introduce a dual-level (i.e., instance-level and cluster-level) contrastive learning auxiliary task to effectively capture different-grained contrastive information within massive unlabeled data. Meanwhile, previous models merely perform the main task and auxiliary tasks in parallel, without considering the relationship among tasks. Therefore, we introduce a hierarchical architecture to explicitly model the correlations between tasks. We conduct extensive experiments across various benchmark datasets, demonstrating that MI-DELIGHT significantly surpasses previous competitive models. It even outperforms popular large language models on several datasets.

The recent Segment Anything Model (SAM) has emerged as a new paradigmatic vision foundation model, showcasing potent zero-shot generalization and flexible prompting. Despite SAM finding applications and adaptations in various domains, its primary limitation lies in the inability to grasp object semantics. In this paper, we present Sambor to seamlessly integrate SAM with the open-vocabulary object detector in an end-to-end framework. While retaining all the remarkable capabilities inherent to SAM, we boost it to detect arbitrary objects from human inputs like category names or reference expressions. Building upon the SAM image encoder, we introduce a novel SideFormer module designed to acquire SAM features adept at perceiving objects and inject comprehensive semantic information for recognition. In addition, we devise an Open-set RPN that leverages SAM proposals to assist in finding potential objects. Consequently, Sambor enables the open-vocabulary detector to equally focus on generalizing both localization and classification sub-tasks. Our approach demonstrates superior zero-shot performance across benchmarks, including COCO and LVIS, proving highly competitive against previous state-of-the-art methods. We aspire for this work to serve as a meaningful endeavor in endowing SAM to recognize diverse object categories and advancing open-vocabulary learning with the support of vision foundation models.

Multimodal large language models (MLLMs) demand considerable computations for inference due to the extensive parameters and the additional input tokens needed for visual information representation. Herein, we introduce Visual Tokens Withdrawal (VTW), a plug-and-play module to boost MLLMs for rapid inference. Our approach is inspired by two intriguing phenomena we have observed: (1) the attention sink phenomenon that is prevalent in LLMs also persists in MLLMs, suggesting that initial tokens and nearest tokens receive the majority of attention, while middle vision tokens garner minimal attention in deep layers; (2) the presence of information migration, which implies that visual information is transferred to subsequent text tokens within the first few layers of MLLMs. As per our findings, we conclude that vision tokens are unnecessary in the deep layers of MLLMs. Thus, we strategically withdraw them at a certain layer, enabling only text tokens to engage in subsequent layers. To pinpoint the ideal layer for VTW, we initially analyze a limited set of tiny datasets and choose the first layer that meets the Kullback-Leibler divergence criterion. Our VTW approach can cut computational overhead by over 40% across diverse multimodal tasks while maintaining performance.

Rainy images suffer from quality degradation due to the synergistic effect of rain streaks and accumulation. The rain streaks are anisotropic and show a specific directional arrangement, while the rain accumulation is isotropic and shows a consistent concentration distribution in local regions. This distribution difference makes unified representation learning for rain streaks and accumulation challenging, which may lead to structure distortion and contrast degradation in the deraining results. To address this problem, a central-surrounding mechanism inspired Synergistic Convolution (SC) is proposed to extract rain streaks and accumulation features simultaneously. Specifically, the SC consists of two parallel novel convolutions: Central-Surrounding Difference Convolution (CSD) and Central-Surrounding Addition Convolution (CSA). In CSD, the difference operation between central and surrounding pixels is injected into the feature extraction process of convolution to perceive the direction distribution of rain streaks. In CSA, the addition operation between central and surrounding pixels is injected into the feature extraction process of convolution to facilitate the modeling of rain accumulation properties. The SC can be used as a general unit to substitute Vanilla Convolution (VC) in current de-raining networks to boost performance. To reduce computational costs, CSA and CSD in SC are merged into a single VC kernel by our parameter equivalent transformation before inferencing. Evaluations of twelve de-raining methods on nine public datasets demonstrate that our proposed SC can comprehensively improve the performance of twelve de-raining networks under various rainy conditions without changing the original network structure or introducing extra computational costs. Even for the current SOTA methods, SC can further achieve SOTA++ performance. The source codes will be publicly available.

Many unsupervised visual anomaly detection methods train an auto-encoder to reconstruct normal samples and then leverage the reconstruction error map to detect and localize the anomalies. However, due to the powerful modeling and generalization ability of neural networks, some anomalies can also be well reconstructed, resulting in unsatisfactory detection and localization accuracy. In this paper, a small coarsely-labeled anomaly dataset is first collected. Then, a coarse-knowledge-aware adversarial learning method is developed to align the distribution of reconstructed features with that of normal features. The alignment can effectively suppress the auto-encoder's reconstruction ability on anomalies and thus improve the detection accuracy. Considering that anomalies often only occupy very small areas in anomalous images, a patch-level adversarial learning strategy is further developed. Although no patch-level anomalous information is available, we rigorously prove that by simply viewing any patch features from anomalous images as anomalies, the proposed knowledge-aware method can also align the distribution of reconstructed patch features with the normal ones. Experimental results on four medical datasets and two industrial datasets demonstrate the effectiveness of our method in improving the detection and localization performance.

Recent research showcases the considerable potential of conditional diffusion models for generating consistent stories. However, current methods, which primarily generate stories in a caption-dependent manner, often overlook the importance of contextual consistency and the relevance of frames during sequential generation. To address this, we propose a novel Rich-contextual Conditional Diffusion Models (RCDMs), a two-stage approach designed to enhance story generation's semantic consistency and temporal consistency. Specifically, in the first stage, the frame-prior transformer diffusion model is presented to predict the frame semantic embedding of the unknown clip by aligning the semantic correlations between the captions and frames of the known clip. The second stage establishes a robust model with rich contextual conditions, including reference images of the known clip, the predicted frame semantic embedding of the unknown clip, and text embeddings of all captions. By jointly injecting these rich contextual conditions at the image and feature levels, RCDMs can generate semantic and temporal consistency stories. Moreover, RCDMs can generate consistent stories with a single forward inference compared to autoregressive models. Our qualitative and quantitative results demonstrate that our proposed RCDMs outperform in challenging scenarios.

Recent advances in differentiable score-based methods for Directed Acyclic Graph (DAG) structure learning have revolutionized the problem of combinatorial structure learning, transforming it into a continuous optimization task. Despite their remarkable success, these methods rely on a key assumption that all samples have the same level of difficulty and no data heterogeneity. When this assumption does not hold, causal discovery algorithms based on it inevitably return networks with many spurious edges. Despite existing research, the current method ignores the reality of outliers in the samples, introducing certain limitations that still result in erroneous edges. Inspired by the rapid decay of the Gaussian distribution as distance from the center increases, we propose an innovative adaptive sample reweighting framework based on asymmetric exponential modulation Gaussian, coined DAG-AEG. DAG-AEG boosts DAG structure learning by analyzing the distribution of sample losses and employing the proposed method for adaptive sample attention. Additionally, it can be adapted to heterogeneous data. We used various causal structure learning methods to test the performance of DAG-AEG on synthetic and real datasets. The experimental results demonstrate that the proposed framework significantly improves the performance across all methods, outperforming existing methods.

In this tool paper, we design, develop, and release BoolXAI, an interpretable machine learning classification approach for Explainable AI (XAI) based on expressive Boolean formulas. The Boolean formula defines a logical rule with tunable complexity according to which input data are classified. Beyond the classical conjunction and disjunction, BoolXAI offers expressive operators such as AtLeast, AtMost, and Choose and their parameterization. This provides higher expressiveness compared to rigid rules- and tree-based approaches. We show how to train BoolXAI classifiers effectively using native local optimization to search the space of feasible formulas. We provide illustrative results on several well-known public benchmarks that demonstrate the competitive nature of our approach compared to existing methods. Our work is embodied in the open-source BoolXAI library with a high-level user interface to serve researchers and practitioners. BoolXAI can be used either as a standalone interpretable classifier or for post-hoc explanations of other black-box models or observed behavior. We highlight several desirable benefits of our tool, especially in industrial settings where rapid experimentation, reusability, reproducibility, deployment, and maintenance are of great interest. Finally, we showcase a deployed service powered by BoolXAI as an enterprise application.

With the widespread use of virtual reality applications, 3D scene generation has become a new challenging research frontier. 3D scenes have highly complex structures and need to ensure that the output is dense, coherent, and contains all necessary structures. Many current 3D scene generation methods rely on pre-trained text-to-image diffusion models and monocular depth estimators. However, the generated scenes occupy large amounts of storage space and often lack effective regularisation methods, leading to geometric distortions. To this end, we propose BloomScene, a lightweight structured 3D Gaussian splatting for crossmodal scene generation, which creates diverse and high-quality 3D scenes from text or image inputs. Specifically, a crossmodal progressive scene generation framework is proposed to generate coherent scenes utilizing incremental point cloud reconstruction and 3D Gaussian splatting. Additionally, we propose a hierarchical depth prior-based regularization mechanism that utilizes multi-level constraints on depth accuracy and smoothness to enhance the realism and continuity of the generated scenes. Ultimately, we propose a structured context-guided compression mechanism that exploits structured hash grids to model the context of unorganized anchor attributes, which significantly eliminates structural redundancy and reduces storage overhead. Comprehensive experiments across multiple scenes demonstrate the significant potential and advantages of our framework compared with several baselines.

Image rescaling (IR) seeks to determine the optimal low-resolution (LR) representation of a high-resolution (HR) image to reconstruct a high-quality super-resolution (SR) image. Typically, HR images with resolutions exceeding 2K possess rich information that is unevenly distributed across the image. Traditional image rescaling methods often fall short because they focus solely on the overall scaling rate, ignoring the varying amounts of information in different parts of the image. To address this limitation, we propose a Block-Based Multi-Scale Image Rescaling Framework (BBMR), tailored for IR tasks involving HR images of 2K resolution and higher. BBMR consists of two main components: the Downscaling Module and the Upscaling Module. In the Downscaling Module, the HR image is segmented into sub-blocks of equal size, with each sub-block receiving a dynamically allocated scaling rate while maintaining a constant overall scaling rate. For the Upscaling Module, we introduce the Joint Super-Resolution method (JointSR), which performs SR on these sub-blocks with varying scaling rates and effectively eliminates blocking artifacts. Experimental results demonstrate that BBMR significantly enhances the SR image quality in the of 2K and 4K test dataset compared to initial network image rescaling methods.

Drug-target interaction prediction (DTI) is essential in various applications including drug discovery and clinical application. There are two perspectives of input data widely used in DTI prediction: Intrinsic data represents how drugs or targets are constructed, and extrinsic data represents how drugs or targets are related to other biological entities. However, any of the two perspectives of input data can be scarce for some drugs or targets, especially for those unpopular or newly discovered. Furthermore, ground-truth labels for specific interaction types can also be scarce. Therefore, we propose the first method to tackle DTI prediction under input data and/or label scarcity. To make our model functional when only one perspective of input data is available, we design two separate experts to process intrinsic and extrinsic data respectively and fuse them adaptively according to different samples. Furthermore, to make the two perspectives complement each other and remedy label scarcity, two experts synergize with each other in a mutually supervised way to exploit the enormous unlabeled data. Extensive experiments on 3 real-world datasets under different extents of input data scarcity and/or label scarcity demonstrate our model outperforms states of the art significantly and steadily, with a maximum improvement of 53.53%. We also test our model without any data scarcity and it still outperforms current methods.

Test-time prompt tuning (TPT) aims to adjust the vision-language models (e.g., CLIP) with learnable prompts during the inference phase. However, previous works overlooked that pre-trained models as a service (MaaS) have become a noticeable trend due to their commercial usage and potential risk of misuse. In the context of MaaS, users can only design prompts in inputs and query the black-box vision-language models through inference APIs, rendering the previous paradigm of utilizing gradient for prompt tuning is infeasible. In this paper, we propose black-box test-time prompt tuning (B²TPT), a novel framework that addresses the challenge of optimizing prompts without gradients in an unsupervised manner. Specifically, B²TPT designs a consistent or confident (CoC) pseudo-labeling strategy to generate high-quality pseudo-labels from the outputs. Subsequently, we propose to optimize low-dimensional intrinsic prompts using a derivative-free evolution algorithm and to project them onto the original text and vision prompts. This strategy addresses the gradient-free challenge while reducing complexity. Extensive experiments across 15 datasets demonstrate the superiority of B²TPT. The results show that B²TPT not only outperforms CLIP's zero-shot inference at test time, but also surpasses other gradient-based TPT methods.

Black-box optimization (BBO) has become increasingly relevant for tackling complex decision-making problems, especially in public policy domains such as police redistricting. However, its broader application in public policymaking is hindered by the complexity of defining feasible regions and the high-dimensionality of decisions. This paper introduces a novel BBO framework, termed as the Conditional And Generative Black-box Optimization (CageBO). This approach leverages a conditional variational autoencoder to learn the distribution of feasible decisions, enabling a two-way mapping between the original decision space and a simplified, constraint-free latent space. The CageBO efficiently handles the implicit constraints often found in public policy applications, allowing for optimization in the latent space while evaluating objectives in the original space. We validate our method through a case study on large-scale police redistricting problems in Atlanta, Georgia. Our results reveal that our CageBO offers notable improvements in performance and efficiency compared to the baselines.

The recent emergence of extreme climate events has significantly raised awareness about sustainable living. In addition to developing energy-saving materials and technologies, existing research mainly relies on traditional methods that encourage behavioral shifts towards sustainability, which can be overly demanding or only passively engaging. In this work, we propose to employ recommendation systems to actively nudge users toward more sustainable choices. We introduce Green Recommender Aligned with Personalized Eating (GRAPE), which is designed to prioritize and recommend sustainable food options that align with users’ evolving preferences. We also design two innovative Green Loss functions that cater to green indicators with either uniform or differentiated priorities, thereby enhancing adaptability across a range of scenarios. Extensive experiments on a real-world dataset demonstrate the effectiveness of our GRAPE.

Generating novel active molecules for a given protein is an extremely challenging task for generative models that requires an understanding of the complex physical interactions between the molecule and its environment. This paper presents a novel generative model, BindGPT, which uses a conceptually simple but powerful approach to create 3D molecules within the protein's binding site. Our model produces molecular graphs and conformations jointly, eliminating the need for an extra graph reconstruction step. We pre-train BindGPT on a large-scale dataset and fine-tune it with reinforcement learning using scores from external simulation software. We demonstrate how a single pre-trained language model can serve at the same time as a 3D molecular generative model, a conformer generator conditioned on the molecular graph, and a pocket-conditioned 3D molecule generator. Notably, the model does not make any representational equivariance assumptions about the domain of generation. We show how such a simple conceptual approach combined with pre-training and scaling can perform on par or better than the current best-specialized diffusion models, language models, and graph neural networks while being two orders of magnitude cheaper to sample.

The Mixture-of-Experts (MoE) structure scales the Transformer-based large language models (LLMs) and improves their performance with only the sub-linear increase in computation resources. Recently, a fine-grained DeepSeekMoE structure is proposed, which can further improve the computing efficiency of MoE without performance degradation. However, the All-to-All communication introduced by MoE has become a bottleneck, especially for the fine-grained structure, which typically involves and activates more experts, hence contributing to heavier communication overhead. In this paper, we propose a novel MoE structure named BigMac, which is also fine-grained but with high communication efficiency. The innovation of BigMac is mainly due to that we abandon the Communicate-Descend-Ascend-Communicate (CDAC) manner used by fine-grained MoE, which leads to the All-to-All communication always taking place at the highest dimension. Instead, BigMac designs an efficient Descend-Communicate-Communicate-Ascend (DCCA) manner. Specifically, we add a descending and ascending projection at the entrance and exit of the expert, respectively, which enables the communication to perform at a very low dimension. Furthermore, to adapt to DCCA, we re-design the structure of small experts, ensuring that the expert in BigMac has enough complexity to address tokens. Experimental results show that BigMac achieves comparable or even better model quality than fine-grained MoEs with the same number of experts and a similar number of total parameters. Equally importantly, BigMac reduces the end-to-end latency by up to 3.09 x for training and increases the throughput by up to 3.11 x for inference on state-of-the-art AI computing frameworks including Megatron, Tutel, and DeepSpeed-Inference.

This paper addresses the challenge of Granularity Competition in fine-grained classification tasks, which arises due to the semantic gap between multi-granularity labels. Existing approaches typically develop independent hierarchy-aware models based on shared features extracted from a common base encoder. However, because coarse-grained levels are inherently easier to learn than finer ones, the base encoder tends to prioritize coarse feature abstractions, which impedes the learning of fine-grained features. To overcome this challenge, we propose a novel framework called the Bidirectional Logits Tree (BiLT) for Granularity Reconcilement. The key idea is to develop classifiers sequentially from the finest to the coarsest granularities, rather than parallelly constructing a set of classifiers based on the same input features. In this setup, the outputs of finer-grained classifiers serve as inputs for coarser-grained ones, facilitating the flow of hierarchical semantic information across different granularities. On top of this, we further introduce an Adaptive Intra-Granularity Difference Learning (AIGDL) approach to uncover subtle semantic differences between classes within the same granularity. Extensive experiments demonstrate the effectiveness of our proposed method.

We present a real-time system that enables bidirectional human-AI learning and teaching in a balancing task that is a realistic analogue of disorientation during piloting and spaceflight. A human subject and autonomous AI model of choice guide each other in maintaining balance using a visual inverted pendulum (VIP) display. We show how AI assistance changes human performance and vice versa.

Considering the ubiquitous phenomenon of missing views in multi-view data, incomplete multi-view learning is a crucial task in many applications. Existing methods usually follow an impute-then-predict strategy for handling this problem. However, they often assume that the view-missing patterns are uniformly random in multi-view data, which does not agree with real-world scenarios. In practice, view-missing patterns often vary across different classes. For example, in the medical field, patients with rare diseases would take more examinations than those with common diseases; in the financial field, high-risk customers tend to receive evaluations from more views than ordinary ones. Hence, we often observe that data-rich classes suffer limited views while data-poor classes suffer limited samples. Previous methods would typically fail due to such biased view-missing patterns. This motivates us to delve into a new biased incomplete multi-view learning problem. To this end, we develop a Reliable Incomplete Multi-view Learning (RIML) method. RIML is a simple yet effective learning-free imputation framework that goes beyond the conventional approaches by considering information from all classes, rather than just relying on individual views or within-class samples. Specifically, we utilize an inter-class association matrix that allows data-poor classes to refer the knowledge from data-rich classes. This enables the construction of more reliable view-specific distributions, from which we perform multiple samplings to recover missing views. Additionally, to obtain a reliable multi-view representation for downstream tasks, we develop an enhanced focal loss with a category-aware marginal term to learn a more distinguishable feature space. Experiments on five multi-view datasets demonstrate that RIML significantly outperforms existing methods in both accuracy and robustness.