Transformer Neural Network Optimized for Real-Time Outlier Detection on FPGAs
A new arXiv preprint (2607.22786) proposes optimizing Transformer neural network inference for real-time anomaly detection in financial time series, such as asset prices. The work addresses the growing volume of financial data that often contains errors or outliers, making downstream processing unreliable. Transformers, known for superior performance in NLP and computer vision, are leveraged for their ability to capture long-range dependencies, which benefits time series modeling and anomaly detection. The optimization targets FPGA deployment for speed and accuracy in data cleaning.
Key facts
- arXiv preprint 2607.22786
- Optimizes Transformer neural network for real-time anomaly detection
- Application to financial time series (asset prices)
- Addresses errors and outliers in financial data
- Data volume increasing, requiring better cleaning methods
- Transformers capture long-range dependencies
- Targets FPGA deployment for inference optimization
Entities
Institutions
- arXiv