REWIND Dataset: Privacy-preserving Speaking Status Segmentation from Multimodal Body Movement Signals in the Wild
Recognizing speaking in humans using machine learning models trained on video and wearable sensor data.
MLAAD: The Multi-Language Audio Anti-Spoofing Dataset
다양한 언어로 구성된 MLAAD 데이터셋의 생성과 활용
Empowering Data Mesh with Federated Learning: Integrating FL into Data Mesh Architecture
Integrating Federated Learning into Data Mesh architecture enhances privacy and decentralized data analysis.
Addressing Concept Shift in Online Time Series Forecasting: Detect-then-Adapt
Novel approach D3A improves model adaptation capability by detecting and adapting to concept drift in online time series forecasting.
Comprehensive Evaluation of GNN Training Systems: A Data Management Perspective
Efficient data management is crucial for successful GNN training, balancing computational and communication loads.
Jellyfish: A Large Language Model for Data Preprocessing and Instruction-Tuning
Instruction-tuned local LLMs enhance DP performance and generalizability.
DS-Agent: Automated Data Science Empowered by Large Language Models and Case-Based Reasoning
DS-Agent utilizes large language models and case-based reasoning to automate data science tasks effectively.
Detection of Unobserved Common Causes in Causal Discovery Using NML Code
The author proposes a method for detecting unobserved common causes in causal discovery using the NML code, extending it to various data types with high performance theoretically and experimentally.
Analyzing the Role of Cosine-Similarity in Embeddings
The author explores how cosine-similarity in embeddings can yield arbitrary and sometimes meaningless results due to the freedom in learned embeddings, cautioning against blind usage and proposing alternatives.
Analyzing Sample Hardness in Data-Centric AI
The author presents a fine-grained taxonomy of hardness types and introduces the Hardness Characterization Analysis Toolkit (H-CAT) to evaluate different Hardness Characterization Methods (HCMs) comprehensively. The goal is to address the lack of consensus and quantitative evaluation in characterizing "hard" samples.