AAAI 2024 论文列表

Existing GAN-based approaches to caricature generation mainly focus on exaggerating a character’s global facial structure. This often leads to the failure in highlighting significant facial features such as big eyes and hook nose. To address this limitation, we propose a new approach termed as G2L-CariGAN, which uses feature maps of spatial dimensions instead of latent codes for geometric exaggeration. G2L-CariGAN first exaggerates the global facial structure of the character on a low-dimensional feature map and then exaggerates its local facial features on a high-dimensional feature map. Moreover, we develop a caricature identity loss function based on feature maps, which well retains the character's identity after exaggeration. Our experiments have demonstrated that G2L-CariGAN outperforms the state-of-arts in terms of the quality of exaggerating a character and retaining its identity.

In this paper, we focus on a realistic yet challenging task, Single Domain Generalization Object Detection (S-DGOD), where only one source domain's data can be used for training object detectors, but have to generalize multiple distinct target domains. In S-DGOD, both high-capacity fitting and generalization abilities are needed due to the task's complexity. Differentiable Neural Architecture Search (NAS) is known for its high capacity for complex data fitting and we propose to leverage Differentiable NAS to solve S-DGOD. However, it may confront severe over-fitting issues due to the feature imbalance phenomenon, where parameters optimized by gradient descent are biased to learn from the easy-to-learn features, which are usually non-causal and spuriously correlated to ground truth labels, such as the features of background in object detection data. Consequently, this leads to serious performance degradation, especially in generalizing to unseen target domains with huge domain gaps between the source domain and target domains. To address this issue, we propose the Generalizable loss (G-loss), which is an OoD-aware objective, preventing NAS from over-fitting by using gradient descent to optimize parameters not only on a subset of easy-to-learn features but also the remaining predictive features for generalization, and the overall framework is named G-NAS. Experimental results on the S-DGOD urban-scene datasets demonstrate that the proposed G-NAS achieves SOTA performance compared to baseline methods. Codes are available at https://github.com/wufan-cse/G-NAS.

It has become a popular paradigm to transfer the knowledge of large-scale pre-trained models to various downstream tasks via fine-tuning the entire model parameters. However, with the growth of model scale and the rising number of downstream tasks, this paradigm inevitably meets the challenges in terms of computation consumption and memory footprint issues. Recently, Parameter-Efficient Fine-Tuning (PEFT) (e.g., Adapter, LoRA, BitFit) shows a promising paradigm to alleviate these concerns by updating only a portion of parameters. Despite these PEFTs having demonstrated satisfactory performance in natural language processing, it remains under-explored for the question: whether these techniques could be transferred to graph-based tasks with Graph Transformer Networks (GTNs)? Therefore, in this paper, we fill this gap by providing extensive benchmarks with traditional PEFTs on a range of graph-based downstream tasks. Our empirical study shows that it is sub-optimal to directly transfer existing PEFTs to graph-based tasks due to the issue of feature distribution shift. To address this issue, we propose a novel structure-aware PEFT approach, named G-Adapter, which leverages graph convolution operation to introduce graph structure information (e.g., graph adjacency matrix) as an inductive bias to guide the updating process. Further, we propose Bregman proximal point optimization to alleviate feature distribution shift by preventing the model from aggressive update. Extensive experiments demonstrate that G-Adapter obtains state-of-the-art performance compared to counterparts on nine graph benchmark datasets based on diverse pre-trained GTNs, and delivers tremendous memory footprint efficiency compared to the conventional paradigm.

Recently Text-to-Video (T2V) synthesis has undergone a breakthrough by training transformers or diffusion models on large-scale datasets. Nevertheless, inferring such large models incurs huge costs. Previous inference acceleration works either require costly retraining or are model-specific. To address this issue, instead of retraining we explore the inference process of two mainstream T2V models using transformers and diffusion models. The exploration reveals the redundancy in temporal attention modules of both models, which are commonly utilized to establish temporal relations among frames. Consequently, we propose a training-free and generalized pruning strategy called F3-Pruning to prune redundant temporal attention weights. Specifically, when aggregate temporal attention values are ranked below a certain ratio, corresponding weights will be pruned. Extensive experiments on three datasets using a classic transformer-based model CogVideo and a typical diffusion-based model Tune-A-Video verify the effectiveness of F3-Pruning in inference acceleration, quality assurance and broad applicability.

Depth uncertainty is a core challenge in 3D human pose estimation, especially when the camera parameters are unknown. Previous methods try to reduce the impact of depth uncertainty by multi-view and/or multi-frame feature fusion to utilize more spatial and temporal information. However, they generally lead to marginal improvements and their performance still cannot match the camera-parameter-required methods. The reason is that their handcrafted fusion schemes cannot fuse the features flexibly, e.g., the multi-view and/or multi-frame features are fused separately. Moreover, the diverse and complicated fusion schemes make the principle for developing effective fusion schemes unclear and also raises an open problem that whether there exist more simple and elegant fusion schemes. To address these issues, this paper proposes an extremely concise unified feature fusion transformer (FusionFormer) with minimized handcrafted design for 3D pose estimation. FusionFormer fuses both the multi-view and multi-frame features in a unified fusion scheme, in which all the features are accessible to each other and thus can be fused flexibly. Experimental results on several mainstream datasets demonstrate that FusionFormer achieves state-of-the-art performance. To our best knowledge, this is the first camera-parameter-free method to outperform the existing camera-parameter-required methods, revealing the tremendous potential of camera-parameter-free models. These impressive experimental results together with our concise feature fusion scheme resolve the above open problem. Another appealing feature of FusionFormer we observe is that benefiting from its effective fusion scheme, we can achieve impressive performance with smaller model size and less FLOPs.

Remote physiology, which involves monitoring vital signs without the need for physical contact, has great potential for various applications. Current remote physiology methods rely only on a single camera or radio frequency (RF) sensor to capture the microscopic signatures from vital movements. However, our study shows that fusing deep RGB and RF features from both sensor streams can further improve performance. Because these multimodal features are defined in distinct dimensions and have varying contextual importance, the main challenge in the fusion process lies in the effective alignment of them and adaptive integration of features under dynamic scenarios. To address this challenge, we propose a novel vital sensing model, named Fusion-Vital, that combines the RGB and RF modalities through the new introduction of pairwise input formats and transformer-based fusion strategies. We also perform comprehensive experiments based on a newly collected and released remote vital dataset comprising synchronized video-RF sensors, showing the superiority of the fusion approach over the previous single-sensor baselines in various aspects.

Programmatic Weak Supervision (PWS) and generative models serve as crucial tools that enable researchers to maximize the utility of existing datasets without resorting to laborious data gathering and manual annotation processes. PWS uses various weak supervision techniques to estimate the underlying class labels of data, while generative models primarily concentrate on sampling from the underlying distribution of the given dataset. Although these methods have the potential to complement each other, they have mostly been studied independently. Recently, WSGAN proposed a mechanism to fuse these two models. Their approach utilizes the discrete latent factors of InfoGAN for the training of the label models and leverages the class-dependent information of the label models to generate images of specific classes. However, the disentangled latent factor learned by the InfoGAN may not necessarily be class specific and hence could potentially affect the label model's accuracy. Moreover, the prediction of the label model is often noisy in nature and can have a detrimental impact on the quality of images generated by GAN. In our work, we address these challenges by (i) implementing a noise-aware classifier using the pseudo labels generated by the label model, (ii) utilizing the prediction of the noise-aware classifier for training the label model as well as generation of class-conditioned images. Additionally, We also investigate the effect of training the classifier with a subset of the dataset within a defined uncertainty budget on pseudo labels. We accomplish this by formalizing the subset selection problem as submodular maximization with a knapsack constraint on the entropy of pseudo labels. We conduct experiments on multiple datasets and demonstrate the efficacy of our methods on several tasks vis-a-vis the current state-of-the-art methods. Our implementation is available at https://github.com/kyrs/subpws-gan

Multivariate Time-Series (MTS) data is crucial in various application fields. With its sequential and multi-source (multiple sensors) properties, MTS data inherently exhibits Spatial-Temporal (ST) dependencies, involving temporal correlations between timestamps and spatial correlations between sensors in each timestamp. To effectively leverage this information, Graph Neural Network-based methods (GNNs) have been widely adopted. However, existing approaches separately capture spatial dependency and temporal dependency and fail to capture the correlations between Different sEnsors at Different Timestamps (DEDT). Overlooking such correlations hinders the comprehensive modelling of ST dependencies within MTS data, thus restricting existing GNNs from learning effective representations. To address this limitation, we propose a novel method called Fully-Connected Spatial-Temporal Graph Neural Network (FC-STGNN), including two key components namely FC graph construction and FC graph convolution. For graph construction, we design a decay graph to connect sensors across all timestamps based on their temporal distances, enabling us to fully model the ST dependencies by considering the correlations between DEDT. Further, we devise FC graph convolution with a moving-pooling GNN layer to effectively capture the ST dependencies for learning effective representations. Extensive experiments show the effectiveness of FC-STGNN on multiple MTS datasets compared to SOTA methods. The code is available at https://github.com/Frank-Wang-oss/FCSTGNN.

The core of pointly-supervised panoptic segmentation is estimating accurate dense pseudo labels from sparse point labels to train the panoptic head. Previous works generate pseudo labels mainly based on hand-crafted rules, such as connecting multiple points into polygon masks, or assigning the label information of labeled pixels to unlabeled pixels based on the artificially defined traversing distance. The accuracy of pseudo labels is limited by the quality of the hand-crafted rules (polygon masks are rough at object contour regions, and the traversing distance error will result in wrong pseudo labels). To overcome the limitation of hand-crafted rules, we estimate pseudo labels with a fully data-driven pseudo label branch, which is optimized by point labels end-to-end and predicts more accurate pseudo labels than previous methods. We also train an auxiliary semantic branch with point labels, it assists the training of the pseudo label branch by transferring semantic segmentation knowledge through shared parameters. Experiments on Pascal VOC and MS COCO demonstrate that our approach is effective and shows state-of-the-art performance compared with related works. Codes are available at https://github.com/BraveGroup/FDD.

Estimating 3D full-body pose from sparse sensor data is a pivotal technique employed for the reconstruction of realistic human motions in Augmented Reality and Virtual Reality. However, translating sparse sensor signals into comprehensive human motion remains a challenge since the sparsely distributed sensors in common VR systems fail to capture the motion of full human body. In this paper, we use well-designed Body Pose Graph (BPG) to represent the human body and translate the challenge into a prediction problem of graph missing nodes. Then, we propose a novel full-body motion reconstruction framework based on BPG. To establish BPG, nodes are initially endowed with features extracted from sparse sensor signals. Features from identifiable joint nodes across diverse sensors are amalgamated and processed from both temporal and spatial perspectives. Temporal dynamics are captured using the Temporal Pyramid Structure, while spatial relations in joint movements inform the spatial attributes. The resultant features serve as the foundational elements of the BPG nodes. To further refine the BPG, node features are updated through a graph neural network that incorporates edge reflecting varying joint relations. Our method's effectiveness is evidenced by the attained state-of-the-art performance, particularly in lower body motion, outperforming other baseline methods. Additionally, an ablation study validates the efficacy of each module in our proposed framework.

Significance testing aims to determine whether a proposition about the population distribution is the truth or not given observations. However, traditional significance testing often needs to derive the distribution of the testing statistic, failing to deal with complex nonlinear relationships. In this paper, we propose to conduct Full Bayesian Significance Testing for neural networks, called nFBST, to overcome the limitation in relationship characterization of traditional approaches. A Bayesian neural network is utilized to fit the nonlinear and multi-dimensional relationships with small errors and avoid hard theoretical derivation by computing the evidence value. Besides, nFBST can test not only global significance but also local and instance-wise significance, which previous testing methods don't focus on. Moreover, nFBST is a general framework that can be extended based on the measures selected, such as Grad-nFBST, LRP-nFBST, DeepLIFT-nFBST, LIME-nFBST. A range of experiments on both simulated and real data are conducted to show the advantages of our method.

Large Language Models (LLM) exhibit zero-shot mathematical reasoning capacity as a behavior emergent with scale, commonly manifesting as chain-of-thoughts (CoT) reasoning. However, multiple empirical findings suggest that this prowess is exclusive to LLMs that have exorbitant sizes (beyond 50 billion parameters). Meanwhile, educational neuroscientists suggest that symbolic algebraic manipulation be introduced around the same time as arithmetic word problems so as to modularize language-to-formulation, symbolic manipulation of the formulation, and endgame arithmetic. In this paper, we start with the hypothesis that much smaller LMs, which are weak at multi-step reasoning, can achieve reasonable arithmetic reasoning if arithmetic word problems are posed as a formalize-then-solve task. In our architecture, which we call SyReLM, the LM serves the role of a translator to map natural language arithmetic questions into a formal language (FL) description. A symbolic solver then evaluates the FL expression to obtain the answer. A small frozen LM, equipped with an efficient low-rank adapter, is capable of generating FL expressions that incorporate natural language descriptions of the arithmetic problem (e.g., variable names and their purposes, formal expressions combining variables, etc.). We adopt policy-gradient reinforcement learning to train the adapted LM, informed by the non-differentiable symbolic solver. This marks a sharp departure from the recent development in tool-augmented LLMs, in which the external tools (e.g., calculator, Web search, etc.) are essentially detached from the learning phase of the LM. SyReLM shows massive improvements (e.g., +30.65 absolute point improvement in accuracy on the SVAMP dataset using GPT-J 6B model) over base LMs, while keeping our testbed easy to diagnose and interpret, and within the reach of most researchers.

The pretrain-finetune paradigm has achieved great success in NLP and 2D image fields because of the high-quality representation ability and transferability of their pretrained models. However, pretraining such a strong model is difficult in the 3D point cloud field due to the limited amount of point cloud sequences. This paper introduces Efficient Point Cloud Learning (EPCL), an effective and efficient point cloud learner for directly training high-quality point cloud models with a frozen CLIP transformer. Our EPCL connects the 2D and 3D modalities by semantically aligning the image features and point cloud features without paired 2D-3D data. Specifically, the input point cloud is divided into a series of local patches, which are converted to token embeddings by the designed point cloud tokenizer. These token embeddings are concatenated with a task token and fed into the frozen CLIP transformer to learn point cloud representation. The intuition is that the proposed point cloud tokenizer projects the input point cloud into a unified token space that is similar to the 2D images. Comprehensive experiments on 3D detection, semantic segmentation, classification and few-shot learning demonstrate that the CLIP transformer can serve as an efficient point cloud encoder and our method achieves promising performance on both indoor and outdoor benchmarks. In particular, performance gains brought by our EPCL are 19.7 AP50 on ScanNet V2 detection, 4.4 mIoU on S3DIS segmentation and 1.2 mIoU on SemanticKITTI segmentation compared to contemporary pretrained models. Code is available at \url{https://github.com/XiaoshuiHuang/EPCL}.

Despite the tremendous success of deep neural networks (DNNs) across various fields, their susceptibility to potential backdoor attacks seriously threatens their application security, particularly in safety-critical or security-sensitive ones. Given this growing threat, there is a pressing need for research into purging backdoors from DNNs. However, prior efforts on erasing backdoor triggers not only failed to withstand increasingly powerful attacks but also resulted in reduced model performance. In this paper, we propose From Toxic to Trustworthy (FTT), an innovative approach to eliminate backdoor triggers while simultaneously enhancing model accuracy. Following the stringent and practical assumption of limited availability of clean data, we introduce a self-attention distillation (SAD) method to remove the backdoor by aligning the shallow and deep parts of the network. Furthermore, we first devise a semi-supervised learning (SSL) method that leverages ubiquitous and available poisoned data to further purify backdoors and improve accuracy. Extensive experiments on various attacks and models have shown that our FTT can reduce the attack success rate from 97% to 1% and improve the accuracy of 4% on average, demonstrating its effectiveness in mitigating backdoor attacks and improving model performance. Compared to state-of-the-art (SOTA) methods, our FTT can reduce the attack success rate by 2 times and improve the accuracy by 5%, shedding light on backdoor cleansing.

The integration of learning and reasoning is one of the key challenges in artificial intelligence and machine learning today. The area of Neuro-Symbolic AI (NeSy) tackles this challenge by integrating symbolic reasoning with neural networks. In our recent work, we provided an introduction to NeSy by drawing several parallels to another field that has a rich tradition in integrating learning and reasoning, namely Statistical Relational Artificial Intelligence (StarAI).

The immense parameter space of Large Language Models (LLMs) endows them with superior knowledge retention capabilities, allowing them to excel in a variety of natural language processing tasks. However, it also instigates difficulties in consistently tuning LMs to incorporate the most recent knowledge, which may further lead LMs to produce inaccurate and fabricated content. To alleviate this issue, we propose a knowledge metabolism framework for LLMs. This framework proactively sustains the credibility of knowledge through an auxiliary external memory component and directly delivers pertinent knowledge for LM inference, thereby suppressing hallucinations caused by obsolete internal knowledge during the LM inference process. Benchmark experiments demonstrate DynaMind's effectiveness in overcoming this challenge. The code and demo of DynaMind are available at: https://github.com/Elfsong/DynaMind.

Retrieving appropriate records from the external knowledge base to generate informative responses is the core capability of end-to-end task-oriented dialogue systems (EToDs). Most of the existing methods additionally train the retrieval model or use the memory network to retrieve the knowledge base, which decouples the knowledge retrieval task from the response generation task, making it difficult to jointly optimize and failing to capture the internal relationship between the two tasks. In this paper, we propose a simple and unified generative model for task-oriented dialogue systems, which recasts the EToDs task as a single sequence generation task and uses maximum likelihood training to train the two tasks in a unified manner. To prevent the generation of non-existent records, we design the prefix trie to constrain the model generation, which ensures consistency between the generated records and the existing records in the knowledge base. Experimental results on three public benchmark datasets demonstrate that our method achieves robust performance on generating system responses and outperforms the baseline systems. To facilitate future research in this area, the code is available at https://github.com/dzy1011/Uni-ToD.

Understanding student behavior in educational settings is critical in improving both the quality of pedagogy and the level of student engagement. While various AI-based models exist for classroom analysis, they tend to specialize in limited tasks and lack generalizability across diverse educational environments. Additionally, these models often fall short in ensuring student privacy and in providing actionable insights accessible to educators. To bridge this gap, we introduce a unified, end-to-end framework by leveraging temporal action detection techniques and advanced large language models for a more nuanced student behavior analysis. Our proposed framework provides an end-to-end pipeline that starts with raw classroom video footage and culminates in the autonomous generation of pedagogical reports. It offers a comprehensive and scalable solution for student behavior analysis. Experimental validation confirms the capability of our framework to accurately identify student behaviors and to produce pedagogically meaningful insights, thereby setting the stage for future AI-assisted educational assessments.

In this paper, we introduce a fresh perspective on the challenges of credit assignment and policy evaluation. First, we delve into the nuances of eligibility traces and explore instances where their updates may result in unexpected credit assignment to preceding states. From this investigation emerges the concept of a novel value function, which we refer to as the ????????????? ????? ????????. Unlike traditional state value functions, bidirectional value functions account for both future expected returns (rewards anticipated from the current state onward) and past expected returns (cumulative rewards from the episode's start to the present). We derive principled update equations to learn this value function and, through experimentation, demonstrate its efficacy in enhancing the process of policy evaluation. In particular, our results indicate that the proposed learning approach can, in certain challenging contexts, perform policy evaluation more rapidly than TD(λ)–a method that learns forward value functions, v^π, ????????. Overall, our findings present a new perspective on eligibility traces and potential advantages associated with the novel value function it inspires, especially for policy evaluation.