Transfer Learning Anomaly Detection Using Source Precision Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing zero-shot deep learning techniques face challenges in efficiently and effectively identifying deep features of the target domain for anomaly detection models using source domain training data, as they typically require large amounts of labeled data and struggle to incorporate the target domain vector into their computations.
Innovation Solution
The approach modifies known deep feature methods by using a first and second source domain precision matrix, along with their respective mean vectors, to compute the anomaly score, allowing the target domain vector to be input as a weight for these computations without changing the input format, enabling efficient anomaly prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional deep learning algorithms are used for anomaly detection, then model performance can be improved with sufficient labeled data, but the requirement for large amounts of labeled training data increases significantly
Solution Approach 1:
The patent changes the computational parameters of the anomaly detection model by modifying how precision matrices and mean vectors are computed and applied. Instead of requiring extensive labeled data for traditional deep learning, the invention uses source domain precision matrices and mean vectors as transferable parameters that can be applied to the target domain with minimal adaptation, thereby reducing the dependency on large amounts of labeled training data while maintaining detection performance
2Quantity of substance
If zero-shot learning techniques are applied to reduce labeled data requirements, then the need for labeled data decreases, but the ability to effectively identify deep features of the target domain is reduced
Solution Approach 1:
The patent introduces source domain precision matrices and mean vectors as intermediary elements that bridge the gap between source and target domains. These intermediaries carry transferable statistical information that helps the model identify deep features in the target domain without requiring extensive labeled target domain data, thus maintaining feature identification accuracy while reducing labeled data requirements
3Quantity of substance
If transfer learning is used to leverage source domain data, then labeled data requirements are reduced, but the complexity of incorporating target domain vectors into computations increases
Solution Approach 1:
The patent inverts the traditional approach by not trying to adapt source domain models to target domain data structures, but rather by computing anomaly scores directly using source domain statistical parameters (precision matrices and mean vectors) in a modified formulation that naturally incorporates target domain information through the anomaly score computation itself, thereby simplifying the transfer learning process
Data Source
AI summary
Embodiments of the invention are directed to a computer system including a memory communicatively coupled to a processor system, where the processor system is operable to perform processor system operations to predict an anomaly in a target domain (TD) dataset. The processor system operations include training a model to perform an anomaly prediction task on a TD. The training includes applying a transfer learning operation that includes learning to predict the anomaly based at least in part on a first source domain (SD) precision matrix computed from a first SD.