Anomaly Detection in Computer Reasoning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems often produce anomalous outcomes without providing clear insights into the training data responsible for these anomalies, making it difficult to correct and remove the faulty data.
Innovation Solution
The approach involves obtaining current context data, determining contextually-determined actions using a reasoning model, identifying the portion of the model causing anomalies, and removing those portions to create a corrected reasoning model, which can then be updated based on subsequent context data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning systems use large amounts of training data to improve accuracy, then the system's predictive capability is improved, but the difficulty of detecting and measuring which data causes anomalies increases
Solution Approach 1:
The patent segments the training data into individual data points with unique identifiers, allowing the system to analyze and identify specific problematic data points rather than treating training data as a monolithic block. This segmentation enables precise anomaly detection by examining the contribution of each individual training example to the model's predictions.
Solution Approach 2:
The patent introduces an intermediary analysis layer that sits between the training data and the final model predictions. This intermediary component tracks and analyzes the relationship between individual training data points and model outputs, enabling the identification of problematic data without requiring direct modification of the core machine learning model.
2Reliability
If machine learning systems use extensive training data to improve performance, then the system's predictive accuracy is improved, but the ability to remove faulty training data becomes more difficult
Solution Approach 1:
The patent performs preliminary analysis during the training phase to identify problematic data points before they cause issues in production. By pre-identifying and flagging anomalous training data through analysis of model predictions and data characteristics, the system prepares for easier correction by having already categorized which data points may need removal or adjustment.
Solution Approach 2:
The patent implements a feedback mechanism where model predictions are continuously analyzed to identify patterns indicating problematic training data. This feedback loop provides information about which training examples are causing anomalies, enabling targeted removal or correction of specific data points while preserving the bulk of the training set.
3Measurement precision
If machine learning systems process large volumes of training data to improve accuracy, then the predictive capability is improved, but the time required to identify and correct anomalies increases
Solution Approach 1:
The patent applies partial analysis by focusing computational resources on identifying only the most problematic portion of training data rather than analyzing every single data point in depth. By using sampling, threshold-based filtering, and prioritization of high-impact data points, the system achieves effective anomaly correction without processing the entire training dataset exhaustively.
Solution Approach 2:
The patent employs techniques to quickly skip through and identify problematic data points using efficient filtering and approximation methods. By implementing fast initial screening followed by targeted detailed analysis only where needed, the system reduces the overall time required to detect and correct anomalies in large training datasets.
Data Source
AI summary
Techniques for detecting and correcting anomalies in computer-based reasoning systems are provided herein. The techniques can include obtaining current context data and determining a contextually-determined action based on the obtained context data and a reasoning model. The reasoning model may have been determined based on multiple sets of training data. The techniques may cause performance of the contextually-determined action and, potentially, receiving an indication that performing the contextually-determined action in the current context resulted in an anomaly. The techniques include determining a portion of the reasoning model that caused the determination of the contextually-determined action based on the obtained context data and causing removal of the portion of the model that caused the determination of the contextually-determined action, to produce a corrected reasoning model. Subsequently, second context data is obtained, a second action is determined based on that data and the corrected reasoning model, and the second contextually-determined action can be performed.