Reasoning Model Anomaly Correction via Training Data Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer-based reasoning systems face difficulties in identifying and correcting anomalies, as it is challenging to determine which training data causes anomalous outcomes, making it hard to remove the problematic data from the model.
Innovation Solution
The techniques involve obtaining current context data, determining a contextually-determined action using a reasoning model, and identifying the portion of the model that caused the anomaly, allowing for the removal of that portion to produce a corrected reasoning model.
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 detection capability is improved, but when anomalies occur it becomes difficult to identify which training data caused the anomaly
Solution Approach 1:
The system implements feedback by monitoring system outcomes and tracing anomalies back to their root causes in the training data. When an anomaly is detected, the system provides feedback information about which specific training data instances contributed to the anomalous behavior, enabling targeted correction of problematic data rather than requiring analysis of all training data.
Solution Approach 2:
The patent introduces an intermediary mechanism that acts as a bridge between the trained machine learning model and the original training data. This intermediary system tracks and records the relationship between training data instances and model decisions, allowing operators to query which specific training data instances influenced anomalous outcomes without manually analyzing the entire training dataset.
2Adaptability or versatility
If the reasoning model uses all available training data to maximize coverage, then the system's versatility is improved, but the system cannot easily remove specific problematic training data instances
Solution Approach 1:
The system segments the training data into individually traceable instances, each tagged and tracked separately. This segmentation allows the system to maintain comprehensive coverage of all training data while enabling selective removal or modification of specific problematic instances. The training data is divided into discrete, manageable units that can be independently analyzed and corrected.
Solution Approach 2:
An intermediary data structure is introduced that maintains the relationship between training data instances and model decisions. This intermediary layer enables efficient querying and selective removal of problematic training data without requiring retraining on the entire dataset, thus maintaining versatility while simplifying correction processes.
3Measurement precision
If the system performs comprehensive analysis of all training data to identify anomalies, then measurement precision is improved, but the time and computational resources required increase significantly
Solution Approach 1:
The system performs preliminary actions by pre-tagging and recording metadata about each training data instance during the training phase. This preliminary organization of data includes information about data sources, characteristics, and relationships, which enables rapid identification of problematic instances when anomalies occur, eliminating the need for comprehensive re-analysis of all training data during correction.
Solution Approach 2:
The system implements efficient feedback mechanisms that provide targeted information about anomalous decisions without requiring full re-analysis. When an anomaly is detected, the feedback system quickly identifies the specific training data instances responsible based on pre-recorded relationships, significantly reducing the time and computational resources needed for correction compared to comprehensive re-analysis.
Data Source
AI summary
Techniques 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 one or more 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.