AAAI 2024 Papers — Page 24
AAAI Conference on Artificial Intelligence · 2331 papers
What Makes Quantization for Large Language Model Hard? An Empirical Study from the Lens of Perturbation
Zhuocheng Gong (Wangxuan Institute of Computer Technology Peking University), Rui Yan (Gaoling School of Artificial Intelligence Renmin University of China)
TransformerLarge Language ModelText
🎯 What it does: Systematically evaluate the quantization effects of LLM from the perspective of perturbation, and propose a non-uniform quantization scheme based on this perspective.
What to Remember: Self-Adaptive Continual Learning for Audio Deepfake Detection
XiaoHui Zhang, Jianhua Tao (Chinese Academy of Sciences)
ClassificationAnomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImageAudio
🎯 What it does: An adaptive continual learning method (Radian Weight Modification, RWM) is proposed, specifically for audio deepfake detection, which can learn new attack types without forgetting old tasks.
When Are Two Lists Better than One?: Benefits and Harms in Joint Decision-Making
Kate Donahue (Cornell University), Kostas Kollias (Google)
🎯 What it does: This paper studies how to select k candidates presented by an algorithm to humans in a human-algorithm collaboration scenario, in order to maximize the probability of selecting the best item.
When CEGAR Meets Regression: A Love Story in Optimal Classical Planning
Martín Pozo (Universidad Carlos III de Madrid), Carlos Linares Lopez (Universidad Carlos III de Madrid)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes the use of regression (backtracking from the goal) in the CEGAR process to generate abstract refinements, thereby achieving reverse abstract refinement in optimal planning.
When Do Program-of-Thought Works for Reasoning?
Zhen Bi (Zhejiang University), Huajun Chen (Zhejiang University)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes the CIRS metric to evaluate the complexity of code reasoning steps, and generates optimal complexity samples through automatic synthesis and hierarchical algorithms to enhance the reasoning ability of LLMs.
When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection
Dongmin Kim (Korea Advanced Institute of Science and Technology), Jaegul Choo (Korea Advanced Institute of Science and Technology)
Anomaly DetectionRecurrent Neural NetworkAuto EncoderTime Series
🎯 What it does: This paper proposes a test-time adaptation method for unsupervised time series anomaly detection based on trend estimation and self-supervised model updates, which can learn new 'normal' patterns in real-time during distribution drift and reduce false positives.
When to Grow? A Fitting Risk-Aware Policy for Layer Growing in Deep Neural Networks
Haihang Wu (University of Melbourne), Saman Halgamuge (University of Melbourne)
ClassificationImage
🎯 What it does: This paper proposes a risk-aware strategy for the timing of deep neural network growth, dynamically adjusting the growth rate to balance overfitting and underfitting;
When to Show a Suggestion? Integrating Human Feedback in AI-Assisted Programming
Hussein Mozannar (Massachusetts Institute of Technology), Eric Horvitz (Microsoft Research)
Recommendation SystemAI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: This study investigates how to utilize programmers' feedback on accepting or rejecting code suggestions to determine when to display or hide AI-generated code suggestions in IDEs, in order to enhance programming efficiency.
Where and How to Attack? A Causality-Inspired Recipe for Generating Counterfactual Adversarial Examples
Ruichu Cai (Guangdong University of Technology), Zhifeng Hao (Shantou University)
Adversarial AttackImage
🎯 What it does: The CADE framework is proposed, which integrates the causal generation process into adversarial attacks to generate more realistic counterfactual adversarial samples, addressing the infeasibility issues caused by traditional attacks that overlook causal relationships.
Which Is More Effective in Label Noise Cleaning, Correction or Filtering?
Gaoxia Jiang (Shanxi University), Deyu Meng (Xi'an Jiaotong University)
ClassificationSupervised Fine-TuningImage
🎯 What it does: This paper theoretically compares two modes of label noise cleaning: correction and filtering, and proposes a fusion cleaning framework (FCF) based on error upper bound analysis.
Who Knows the Answer? Finding the Best Model and Prompt for Each Query Using Confidence-Based Search
Walter Gerych (Massachusetts Institute of Technology), Praveen Venkateswaran (International Business Machines)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the PICS method, which selects the most suitable LLM and prompt template at each input level by predicting confidence, and uses confidence sampling to obtain the most reliable answer in the final return.
Wikiformer: Pre-training with Structured Information of Wikipedia for Ad-Hoc Retrieval
Weihang Su (Tsinghua University), Shengluan Hou (Tsinghua University)
RetrievalTransformerContrastive LearningText
🎯 What it does: Designed the Wikiformer pre-training framework, utilizing the structural information of Wikipedia to construct four self-supervised pre-training objectives for retrieval tasks, enhancing the effectiveness of document relevance learning.
WikiSQE: A Large-Scale Dataset for Sentence Quality Estimation in Wikipedia
Kenichiro Ando (RIKEN AIP), Mamoru Komachi (Hitotsubashi University)
ClassificationTransformerLarge Language ModelText
🎯 What it does: A WikiSQE dataset was constructed to evaluate the quality of Wikipedia sentences;
Winnie: Task-Oriented Dialog System with Structure-Aware Contrastive Learning and Enhanced Policy Planning
Kaizhi Gao (Beijing Institute of Technology), Suli Zou (Beijing Institute of Technology)
TransformerContrastive LearningText
🎯 What it does: This paper proposes an end-to-end task-oriented dialogue system named Winnie, which enhances dialogue quality by utilizing dialogue structure information and system action classification.
Working Memory Capacity of ChatGPT: An Empirical Study
Dongyu Gong (Yale University), Dingmin Wang (University of Oxford)
TransformerLarge Language ModelText
🎯 What it does: This paper systematically evaluates the working memory capacity of ChatGPT in Verbal and Spatial n-back tasks, finding its upper limit to be about 3 levels, similar to humans;
Worst-Case VCG Redistribution Mechanism Design Based on the Lottery Ticket Hypothesis
Mingyu Guo (University of Adelaide)
OptimizationReinforcement Learning
🎯 What it does: A worst-case VCG redistribution mechanism is designed for public project problems, maximizing the worst-case allocation efficiency ratio by learning a differentiable payment function h.
X-RefSeg3D: Enhancing Referring 3D Instance Segmentation via Structured Cross-Modal Graph Neural Networks
Zhipeng Qian (Xiamen University), Xiaoshuai Sun (Xiamen University)
Object DetectionSegmentationGraph Neural NetworkTextPoint Cloud
🎯 What it does: This paper proposes X-RefSeg3D, an end-to-end model for reference-based 3D instance segmentation that constructs cross-modal scene graphs and utilizes structured graph neural networks for entity perception fusion and relationship-driven interaction.
X4D-SceneFormer: Enhanced Scene Understanding on 4D Point Cloud Videos through Cross-Modal Knowledge Transfer
Linglin Jing (Shanghai AI laboratory), Zhen Li (CUHK-Shenzhen)
RecognitionSegmentationTransformerContrastive LearningVideoPoint Cloud
🎯 What it does: By introducing RGB sequences during the training phase and employing the cross-modal knowledge transfer framework X4D-SceneFormer, the understanding capability of 4D point cloud videos (which include the temporal dimension) is enhanced, allowing tasks to be completed using only point cloud data during inference.
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
Zhouhong Gu (Fudan University), Yanghua Xiao (Fudan University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the Xiezhi evaluation benchmark for assessing the interdisciplinary knowledge understanding of large language models.
XKD: Cross-Modal Knowledge Distillation with Domain Alignment for Video Representation Learning
Pritam Sarkar (Queen's University), Ali Etemad (Queen's University)
ClassificationRecognitionKnowledge DistillationRepresentation LearningTransformerAuto EncoderVideoMultimodalityAudio
🎯 What it does: This paper proposes a self-supervised framework XKD, which combines mask reconstruction and cross-modal knowledge distillation to learn a unified representation of audio and video, and incorporates a domain alignment strategy to facilitate effective information sharing between the two modalities.
You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification
Yongqiang Zheng (Guangdong University of Foreign Studies), Xia Li (Guangdong University of Foreign Studies)
ClassificationGraph Neural NetworkContrastive LearningText
🎯 What it does: A method for one-time reading of sentences and simultaneously predicting sentiment polarity for all aspects (One-to-Many ABSA) is proposed, called YORO;
Your Career Path Matters in Person-Job Fit
Zhuocheng Gong (Peking University), Rui Yan (Renmin University of China)
Recommendation SystemRecurrent Neural NetworkContrastive LearningText
🎯 What it does: This study investigates a job seeker and position matching framework based on career path preferences (consistency, similarity, continuity) called WEPJM, which integrates career path information into the job matching scoring using a multi-task learning approach.
YTCommentQA: Video Question Answerability in Instructional Videos
Saelyne Yang (Korea Advanced Institute of Science and Technology), Moontae Lee (LG AI Research)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: A dataset named YTCommentQA has been constructed, containing real user questions from YouTube tutorial videos, with annotations indicating whether the questions can be answered in the video and the required modality (visual, script, or both).
z-SignFedAvg: A Unified Stochastic Sign-Based Compression for Federated Learning
Zhiwei Tang (Chinese University of Hong Kong), Tsung-Hui Chang (Chinese University of Hong Kong)
OptimizationFederated LearningImage
🎯 What it does: This paper designs a random symbol compression method with noise perturbation and proposes the first Federated Averaging algorithm, z-SignFedAvg, that supports multiple local SGD, significantly reducing communication overhead.
Zero-1-to-3: Domain-Level Zero-Shot Cognitive Diagnosis via One Batch of Early-Bird Students towards Three Diagnostic Objectives
Weibo Gao (University of Science and Technology of China), Yuanjing He (Open University of China)
Domain AdaptationTransformerLarge Language ModelTabular
🎯 What it does: Proposes the Zero-1-to-3 framework, which utilizes a single early student batch to achieve cross-domain zero-shot cognitive diagnosis, balancing diagnostic accuracy, signal propagation, and domain adaptation.
Zero-Shot Aerial Object Detection with Visual Description Regularization
Zhengqing Zang (Sichuan University), Jiancheng Lv (Sichuan University)
Object DetectionVision Language ModelImageText
🎯 What it does: A zero-shot aerial target detection method called DescReg is proposed, which utilizes textual descriptions for visual regularization.
Zero-Shot Task Adaptation with Relevant Feature Information
Atsutoshi Kumagai (NTT Computer and Data Science Laboratories), Yasuhiro Fujiwara (NTT Communication Science Laboratories)
ClassificationDomain AdaptationMeta LearningImageText
🎯 What it does: A zero-shot task adaptation method based on meta-learning is proposed, utilizing a small number of relevant features to learn a classifier for the target task.
Zero-Sum Games between Mean-Field Teams: Reachability-Based Analysis under Mean-Field Sharing
Yue Guan (Georgia Institute of Technology), Panagiotis Tsiotras (Georgia Institute of Technology)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper proposes a framework for discrete zero-sum mean field team games (ZS-MFTG) to describe the situation where two large teams compete at the team level while agents within the teams cooperate. It provides a dynamic programming solution method simplified by consistent strategies in the limit of infinite crowds, ultimately achieving ϵ-optimality in the original finite crowd game.
Zhongjing: Enhancing the Chinese Medical Capabilities of Large Language Model through Expert Feedback and Real-World Multi-Turn Dialogue
Songhua Yang (Zhengzhou University), Hongying Zan (Zhengzhou University)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical Data
🎯 What it does: Developed the first end-to-end trained Chinese medical large language model Zhongjing, and constructed a dataset of 70,000 real doctor-patient multi-turn dialogues called CMtMedQA;
ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-Order Optimization
Shuoran Jiang (Harbin Institute of Technology), Xiaobao Song (Institute of Data Security)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a zero-order optimization-based adaptive momentum and uncertainty adjustment operator (ZO-AdaMU), which improves gradient estimation by placing momentum on simulated perturbations, thereby reducing oscillation, alleviating overfitting, and enhancing convergence speed.
ZOOM: Learning Video Mirror Detection with Extremely-Weak Supervision
Ke Xu (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: The ZOOM method is proposed, which utilizes the extremely weak 0/1 mirror presence indicator to train a mirror detector, enabling mirror localization and segmentation without the need for pixel-level annotations.