Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy

arXiv — cs.LG•Wednesday, November 12, 2025 at 5:00:00 AM

The introduction of TimeRCD marks a significant advancement in time series anomaly detection (TSAD), a critical area in artificial intelligence. Traditional models often struggle with subtle anomalies due to their reliance on reconstruction-based objectives, which can misinterpret complex normal patterns, leading to high false positive and negative rates. TimeRCD overcomes these challenges by employing a new pre-training paradigm known as Relative Context Discrepancy (RCD). This model is designed to explicitly detect anomalies by identifying significant discrepancies between adjacent time windows, rather than reconstructing inputs. Implemented with a standard Transformer architecture, TimeRCD captures contextual shifts that previous methods often miss. Extensive experiments have demonstrated that TimeRCD significantly outperforms existing general-purpose and anomaly-specific models in zero-shot TSAD across diverse datasets, showcasing its potential to improve the reliability of anomaly…

— via World Pulse Now AI Editorial System

Read Original

Was this article worth reading? Share it

Recommended Readings

arXiv — cs.LGa day ago

ChemFixer: Correcting Invalid Molecules to Unlock Previously Unseen Chemical Space

PositiveArtificial Intelligence

ChemFixer is a new framework designed to correct invalid molecules generated by deep learning-based molecular generation models. These models have shown promise in exploring chemical spaces for potential drug candidates, but often produce chemically invalid outputs. ChemFixer utilizes a transformer architecture and is fine-tuned on a dataset of valid and invalid molecular pairs. Evaluations indicate that it enhances molecular validity while maintaining the chemical and biological properties of the original outputs, thus expanding the usable chemical space.

Read full article

via arXiv — cs.LG

arXiv — stat.ML2 days ago

Likelihood-guided Regularization in Attention Based Models

PositiveArtificial Intelligence

The paper introduces a novel likelihood-guided variational Ising-based regularization framework for Vision Transformers (ViTs), aimed at enhancing model generalization while dynamically pruning redundant parameters. This approach utilizes Bayesian sparsification techniques to impose structured sparsity on model weights, allowing for adaptive architecture search during training. Unlike traditional dropout methods, this framework learns task-adaptive regularization, improving efficiency and interpretability in classification tasks involving structured and high-dimensional data.

Read full article

via arXiv — stat.ML

arXiv — cs.LG3 days ago

Unitho: A Unified Multi-Task Framework for Computational Lithography

PositiveArtificial Intelligence

Unitho is a unified multi-task framework designed for computational lithography, leveraging the Transformer architecture. It addresses critical tasks such as mask generation, rule violation detection, and layout optimization, which have traditionally been performed in isolation due to limited datasets. Trained on a large-scale industrial lithography simulation dataset comprising hundreds of thousands of cases, Unitho demonstrates significant effectiveness and generalizability, outperforming academic baselines in experimental results.

Read full article

via arXiv — cs.LG

arXiv — cs.CL3 days ago

Equilibrium Dynamics and Mitigation of Gender Bias in Synthetically Generated Data

NeutralArtificial Intelligence

The paper titled 'Equilibrium Dynamics and Mitigation of Gender Bias in Synthetically Generated Data' explores the dynamics of gender bias in synthetic datasets generated through recursive prompting with large language models. It examines three initial bias levels (0.1, 0.3, 0.6) and evaluates four mitigation strategies. The findings indicate that low initial bias amplifies by 36%, while high initial bias decays by 26%. Notably, contrastive augmentation significantly reduces downstream bias by 98.8% for low initial bias and 91% on average, despite higher embedding-based bias scores.

Read full article

via arXiv — cs.CL