AAAI 2025 Papers

Autonomous robots are becoming more versatile and widespread in our daily lives. From autonomous vehicles to companion robots for senior care, these human-centric systems must demonstrate a high degree of reliability in order to build trust and, ultimately, deliver social value. How safe is safe enough for robots to be wholeheartedly trusted by society? Is it sufficient if an autonomous vehicle can avoid hitting a fallen cyclist 99.9% of the time? What if this rate can only be achieved by the vehicle always stopping and waiting for the human to move out of the way? I argue that, for trustworthy deployment of robots in human-populated space, we need to complement standard statistical methods with clear-cut robust safety assurances under a vetted set of operation conditions. We need runtime learning to minimize the robot’s performance loss during safety-enforcing maneuvers by reducing its inherent uncertainty induced by its human peers, for example, their intent (does a human driver want to merge, cut behind, or stay in the lane?) or response (if the robot comes closer, how will the human react?). We need to close the loop between the robot’s learning and decision-making so that it can optimize efficiency by anticipating how its ongoing interaction with the human may affect the evolving uncertainty, and ultimately, its long-term performance.

Children's mental health is crucial for their development, but it's often overlooked, leading to psychological issues. Many children struggle to express their thoughts and feelings effectively. To address this issue, we have proposed a novel approach to analyze children's drawings for psychological screening using artificial intelligence. Specifically, we're focusing on the `draw a person' (DAP) test, where a child's drawing is used to identify potential indicators of their mental and emotional state. Thus, we are introducing an AI-powered technique to automate the psychological screening process for children using the DAP test, which a human professional would traditionally conduct. The screening tool would suggest whether the child needs or doesn’t need further psychological referral. We have collected a dataset consisting of children's drawings and labeled them by experts as either `need' or `no need', indicating whether the child needs or does not need a referral. We have proposed two alternative approaches for the screening process. The first approach consists of extracting features from the drawings following expert guidelines and training a classification model using the features to classify the drawing as either `need' or `no need'. We also propose an out-of-the-box technique applying prompt engineering on state-of-the-art LLMs to automatically extract features from the images. The second approach involves training an image classification model using the drawings. Both approaches are challenged by the issue of class imbalance, as most of the drawings correspond to the `no-need' class. To address this challenge, we introduce Siamese++, a novel Siamese network for image classification, which uses feature embedding and an adaptive distance threshold for classification, instead of the nearest neighbor classification employed by traditional Siamese. Our proposed method achieves a high F1 score (up to 88%) even with a large class imbalance and without the need for any image augmentation. Thus, we have proposed an innovative interdisciplinary integration of AI with psychology and developed novel techniques to solve the real-world problem of psychological screening.

In the rapidly advancing field of AI-assisted medical diagnosis, the generation of medical reports for Chest X-rays (CXR) has significantly improved with the increased availability of radiographs and their corresponding reports. However, these reports often contain complex medical terminology, making them difficult for patients and non-healthcare professionals to understand. In this study, we introduce a strategy called Chained Prompting for Improved Readability of Medical Reports (CPIR-MR), which translates original medical reports into more comprehensible language. Our primary contribution is the creation of a new extension to the IU X-Ray dataset, providing Simplified Medical Reports (SMRs) generated by CPIR-MR. Additionally, we demonstrate that standard methodologies can effectively produce these simplified reports by proposing a multi-modal text decoder (MTD) that combines BLIP with a classification network to generate simplified medical explanations (SMEs) when fine-tuned on SMRs.

Cross-view geo-localization aims at determining the geographic location of a query image by matching the reference images. The matching pairs can be captured from diverse perspectives, such as those from satellites and drones. Most existing methods are supervised that require input of location-labeled images or matched and unmatched image pairs for training, resulting in high labor costs. Moreover, current unsupervised methods perform instances matching directly between different perspectives with dramatic discrepancies, resulting in poor performance. To address these issues, this paper proposes a novel matching and alignment framework from coarse instance-cluster level to fine intermediate instance level for unsupervised cross-view geo-localization. We first introduces cluster-based contrastive learning, assigning pseudo-labels to the instances and generate clusters within each view. Then we design a cross-view location alignment module that fully exploits the feature relationships between instances and clusters for intra- and inter-views. Finally, we design an intermediate state transition module that facilitates further alignment between views by constructing intermediate states and bringing both views closer to the intermediate domain simultaneously. Extensive experiments demonstrate that our method surpasses state-of-the-art unsupervised cross-view geo-localization methods and even achieves comparable performance to state-of-the-art supervised methods.

In this paper, we present CAD2Program, a new method for reconstructing 3D parametric models from 2D CAD drawings. Our proposed method is inspired by recent successes in vision-language models (VLMs), and departs from traditional methods which rely on task-specific data representations and/or algorithms. Specifically, on the input side, we simply treat the 2D CAD drawing as a raster image, regardless of its original format, and encode the image with a standard ViT model. We show that such an encoding scheme achieves competitive performance against existing methods that operate on vector-graphics inputs, while imposing substantially fewer restrictions on the 2D drawings. On the output side, our method auto-regressively predicts a general-purpose language describing 3D parametric models in text form. Compared to other sequence modeling methods for CAD which use domain-specific sequence representations with fixed-size slots, our text-based representation is more flexible, and can be easily extended to arbitrary geometric entities and semantic or functional properties. Experimental results on a large-scale dataset of cabinet models demonstrate the effectiveness of our method.

Video question answering (VideoQA) aims to answer natural language questions according to the given videos. Although existing models perform well in the factoid VideoQA task, they still face challenges in deep video understanding (DVU) task, which focuses on story videos. Compared to factoid videos, the most significant feature of story videos is storylines, which are composed of complex interactions and long-range evolvement of core story topics including characters, actions and locations. Understanding these topics requires models to possess DVU capability. However, existing DVU datasets rarely organize questions according to these story topics, making them difficult to comprehensively assess VideoQA models' DVU capability of complex storylines. Additionally, the question quantity and video length of these dataset are limited by high labor costs of handcrafted dataset building method. In this paper, we devise a large language model based multi-agent collaboration framework, StoryMind, to automatically generate a new large-scale DVU dataset. The dataset, FriendsQA, derived from the renowned sitcom Friends with an average episode length of 1,358 seconds, contains 44.6K questions evenly distributed across 14 fine-grained topics. Finally, We conduct comprehensive experiments on 10 state-of-the-art VideoQA models using the FriendsQA dataset.

Multi-modal multi-party conversation (MMC) is a less studied yet important topic of research due to that it well fits real-world scenarios and thus potentially has more widely-used applications. Compared with the traditional multi-modal conversations, MMC requires stronger character-centered understanding abilities as there are many interlocutors appearing in both the visual and textual context. To facilitate the study of this problem, we present Friends-MMC in this paper, an MMC dataset that contains 24,000+ unique utterances paired with video context. To explore the character-centered understanding of the dialogue, we also annotate the speaker of each utterance, the names and bounding bboxes of faces that appear in the video. Based on this Friends-MMC dataset, we further study two fundamental MMC tasks: conversation speaker identification and conversation response prediction, both of which have the multi-party nature with the video or image as visual context. For conversation speaker identification, we demonstrate the inefficiencies of existing methods such as pre-trained models, and propose a simple yet effective baseline method that leverages an optimization solver to utilize the context of two modalities to achieve better performance. For conversation response prediction, we fine-tune generative dialogue models on Friend-MMC, and analyze the benefits of speaker information. The code and dataset will be publicly available, and thus we call for more attention on modelling speaker information when understanding conversations.

Contrastive learning underpins most current self-supervised time series representation methods. The strategy for constructing positive and negative sample pairs significantly affects the final representation quality. However, due to the continuous nature of time series semantics, the modeling approach of contrastive learning struggles to accommodate the characteristics of time series data. This results in issues such as difficulties in constructing hard negative samples and the potential introduction of inappropriate biases during positive sample construction. Although some recent works have developed several scientific strategies for constructing positive and negative sample pairs with improved effectiveness, they remain constrained by the contrastive learning framework. To fundamentally overcome the limitations of contrastive learning, this paper introduces Frequency-masked Embedding Inference (FEI), a novel non-contrastive method that completely eliminates the need for positive and negative samples. The proposed FEI constructs 2 inference branches based on a prompting strategy: 1) Using frequency masking as prompts to infer the embedding representation of the target series with missing frequency bands in the embedding space, and 2) Using the target series as prompts to infer its frequency masking embedding. In this way, FEI enables continuous semantic relationship modeling for time series. Experiments on 8 widely used time series datasets for classification and regression tasks, using linear evaluation and end-to-end fine-tuning, show that FEI significantly outperforms existing contrastive-based methods in terms of generalization. This study provides new insights into self-supervised representation learning for time series.

By adaptively controlling the density and generating more Gaussians in regions with high-frequency information, 3D Gaussian Splatting (3DGS) can better represent scene details. From the signal processing perspective, representing details usually needs more Gaussians with relatively smaller scales. However, 3DGS currently lacks an explicit constraint linking the density and scale of 3D Gaussians across the domain, leading to 3DGS using improper-scale Gaussians to express frequency information, resulting in the loss of accuracy. In this paper, we propose to establish a direct relation between density and scale through the reparameterization of the scaling parameters and ensure the consistency between them via explicit constraints (i.e., density responds well to changes in frequency). Furthermore, we develop a frequency-aware density control strategy, consisting of densification and deletion, to improve representation quality with fewer Gaussians. A dynamic threshold encourages densification in high-frequency regions, while a scale-based filter deletes Gaussians with improper scale. Experimental results on various datasets demonstrate that our method outperforms existing state-of-the-art methods quantitatively and qualitatively.

In this paper, we propose FreqTS, a novel Frequency-Aware Token Selection approach for accelerating diffusion models without requiring retraining. Diffusion models have gained significant attention in the field of image synthesis due to their impressive generative capabilities. However, these models often suffer from high computational costs, primarily due to the sequential denoising process and large model size. Additionally, diffusion models tend to prioritize low-frequency features, leading to sub-optimal quantitative results. To address these challenges, FreqTS introduces an amplitude-based sorting method that separates Token features in the frequency domain of diffusion models into high-frequency and low-frequency subsets. It then utilizes fast Token Selection to reduce the presence of low-frequency features, effectively reducing the computational overhead. Moreover, FreqTS incorporates a Bayesian hyper-parameter search to dynamically assign different selection strategies for various denoising processes. Extensive experiments conducted on Stable Diffusion series models, PixArt-Alpha, LCM, and other models demonstrate that FreqTS achieves a minimum acceleration of 2.3× without the need for retraining. Furthermore, FreqTS showcases its versatility by being applicable to different sampling techniques and compatible with other dimension-specific acceleration algorithms.

In the pursuit of efficient vision architectures, substantial efforts have been devoted to optimizing operator efficiency. Depth-wise separable operators, such as DWConv, are found cheap in both FLOPs and parameters. As a result, they are increasingly incorporated into efficient backbones, trading for deeper and wider architectures to enhance performance. However, separable operators are not really fast on devices due to the discontinuous memory access requirements. In this paper, we propose FreeNets, a family of simple and efficient backbones that free the separable operation to further accelerate the running speed. We introduce sparse sampling mixers (S2-Mixer) to supersede existing separable token mixers. The S2-Mixer samples multiple segments of partially continuous signals across spatial and channel dimensions for convolutional processing, achieving extremely fast on-device speed. The sparse sampling also enables S2-Mixer to capture long-range pixel relationships from dynamic receptive fields. Furthermore, we introduce a Shift Feed-Forward Network (ShiftFFN) as a faster alternative to existing channel mixers. It utilizes a shift neck architecture that aggregates global information to shift features, enabling faster channel mixing while incorporating global pixel information. Extensive experiments demonstrate that FreeNet offers a superior accuracy-efficiency tradeoff compared to the latest efficient models. On ImageNet-1k, FreeNet-S2 outperforms the StarNet-S4 by 0.4% in top-1 accuracy, while running around 40% faster on desktop GPU and 15% faster on Mobile GPU.

Text-to-video diffusion models have made remarkable advancements. Driven by their ability to generate temporally coherent videos, research on zero-shot video editing using these fundamental models has expanded rapidly. To enhance editing quality, structural controls are frequently employed in video editing. Among these techniques, cross-attention mask control stands out for its effectiveness and efficiency. However, when cross-attention masks are naively applied to video editing, they can introduce artifacts such as blurring and flickering. Our experiments uncover a critical factor overlooked in previous video editing research: cross-attention masks are not consistently clear but vary with model structure and denoising timestep. To address this issue, we propose the metric Mask Matching Cost (MMC) that quantifies this variability and propose FreeMask, a method for selecting optimal masks tailored to specific video editing tasks. Using MMC-selected masks, we further improve the masked fusion mechanism within comprehensive attention features, e.g., temp, cross, and self-attention modules. Our approach can be seamlessly integrated into existing zero-shot video editing frameworks with better performance, requiring no control assistance or parameter fine-tuning but enabling adaptive decoupling of unedited semantic layouts with mask precision control. Extensive experiments demonstrate that FreeMask achieves superior semantic fidelity, temporal consistency, and editing quality compared to state-of-the-art methods.

Text-to-image diffusion model has inspired research into text-to-data synthesis without human intervention, where spatial attentions correlated with semantic entities in text prompts are primarily interpreted as pseudo-masks. However, these vannila attentions often deliver visual-linguistic discrepancies, in which the associations between image features and entity-level tokens are unstable and divergent, yielding inferior masks for realistic applications, especially in more practical open-vocabulary settings. To tackle this issue, we propose a novel text-guided self-driven generative paradigm, termed FreeGen, which addresses the discrepancies by recalibrating intrinsic visual-linguistic correlations and serves as a real-data-free method to automatically synthesize open-vocabulary pixel-level data for arbitrary entities. Specifically, we first learn an Attention Self-Rectification mechanism to reproject the inherent attention matrices to achieve robust semantic alignment, thereby obtaining class-discriminative masks. A Temporal Fluctuation Factor is present to assess mask quality based on its variation over uniform sampling timesteps, enabling the selection of reliable masks. These masks are then employed as self-supervised signals to support the learning of an Entity-level Grounding Decoder in a self-training manner, thus producing open-vocabulary segmentation results. Extensive experiments show that the existing segmenters trained on FreeGen narrow the performance gap with real data counterparts and remarkably outperform the state-of-the-art methods.

We propose a novel hybrid calibration-free method FreeCap to accurately capture global multi-person motions in open environments. Our system combines a single LiDAR with expandable moving cameras, allowing for flexible and precise motion estimation in a unified world coordinate. In particular, We introduce a local-to-global pose-aware cross-sensor human-matching module that predicts the alignment among each sensor, even in the absence of calibration. Additionally, our coarse-to-fine sensor-expandable pose optimizer further optimizes the 3D human key points and the alignments, it is also capable of incorporating additional cameras to enhance accuracy. Extensive experiments on Human-M3 and FreeMotion datasets demonstrate that our method significantly outperforms state-of-the-art single-modal methods, offering an expandable and efficient solution for multi-person motion capture across various applications.

We propose an approach for reconstructing free-moving object from a monocular RGB video. Most existing methods either assume scene prior, hand pose prior, object category pose prior, or rely on local optimization with multiple sequence segments. We propose a method that allows free interaction with the object in front of a moving camera without relying on any prior, and optimizes the sequence globally without any segments. We progressively optimize the object shape and pose simultaneously based on an implicit neural representation. A key aspect of our method is a virtual camera system that reduces the search space of the optimization significantly. We evaluate our method on the standard HO3D dataset and a collection of egocentric RGB sequences captured with a head-mounted device. We demonstrate that our approach outperforms most methods significantly, and is on par with recent techniques that assume prior information.

Tree ensembles, including boosting methods, are highly effective and widely used for tabular data. However, large ensembles lack interpretability and require longer inference times. We introduce a method to prune a tree ensemble into a reduced version that is "functionally identical" to the original model. In other words, our method guarantees that the prediction function stays unchanged for any possible input. As a consequence, this pruning algorithm is lossless for any aggregated metric. We formalize the problem of functionally identical pruning on ensembles, introduce an exact optimization model, and provide a fast yet highly effective method to prune large ensembles. Our algorithm iteratively prunes considering a finite set of points, which is incrementally augmented using an adversarial model. In multiple computational experiments, we show that our approach provides a "free lunch", significantly reducing the ensemble size without altering the model's behavior. Thus, we can preserve state-of-the-art performance at a fraction of the original model's size.

In this paper, we propose a novel Temporal Sequence-Aware-Model (TSAM) for few-shot action recognition (FSAR), which incorporates a sequential perceiver adapter into the pre-training framework, to integrate both the spatial information and the sequential temporal dynamics into the feature embeddings. Different from the existing fine-tuning approaches that capture temporal information by exploring the relationships among all the frames, our perceiver-based adapter recurrently captures the sequential dynamics alongside the timeline, which could perceive the frame order change. To obtain the discriminative representations for each class, we extend a textual corpus for each class derived from the large language models (LLMs) and enrich the visual prototypes by integrating the contextual semantic information. Besides, We introduce an unbalanced optimal transport strategy for feature matching that mitigates the impact of class-unrelated features, thereby facilitating more effective decision-making. Experimental results on five FSAR datasets demonstrate that our method establishes a new benchmark, outperforming the second-best competitors.

Visual Reinforcement Learning (RL) facilitates learning directly from raw images; however, the domain gap between training and testing environments frequently leads to a decline in performance within unseen environments. In this paper, we propose Fourier Guided Adaptive Adversarial Augmentation (FGA3), a novel augmentation method that maintains semantic consistency. We focus on style augmentation in the frequency domain by keeping the phase and altering the amplitude to preserve the state of the original data. For adaptive adversarial perturbation, we reformulate the worst-case problem to RL by employing adversarial example training, which leverages value loss and cosine similarity within a semantic space. Moreover, our findings illustrate that cosine similarity is effective in quantifying feature distances within a semantic space. Extensive experiments on DMControl-GB and Procgen have shown that FGA3 is compatible with a wide range of visual RL algorithms, both off-policy and on-policy, and significantly improves the robustness of the agent in unseen environments.

Recent years have witnessed tremendous successes of learning for sequential decision-making, and in particular, Reinforcement Learning (RL). Prominent application examples include playing Go and video games, robotics, autonomous driving, and recently large language models. Most such success stories naturally involve "multi-agents". Hence, there has been surging research interest in advancing Multi-Agent Learning in Dynamic Environments, particularly, multi-agent RL (MARL), to which my research has led and made significant contributions. My work has established both sample and computational complexities of learning in Stochastic Games, the most fundamental model of MARL, and advocated a unique Economics perspective of independent learning in Stochastic Games. My work has also initiated the recent studies of distributed and networked MARL, with applications in robust adversarial RL, offline RL, and Robotics. This paper will survey my notable contributions along this journey of developing the foundations of multi-agent learning in dynamic environments.