ML Model Exit Control for Atypical Wireless Input Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ML models in wireless communications systems struggle to effectively determine when to stop using AI/ML-based algorithms, particularly when encountering atypical data, leading to potential inefficiencies and suboptimal performance.
Innovation Solution
Implement systems and techniques to measure the similarity and/or dissimilarity between input data and training data, comparing the results to a threshold value to decide whether to continue using the ML model, potentially switching to non-ML based algorithms or different ML models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ML models are continuously used without monitoring data similarity, then system simplicity is maintained, but system reliability deteriorates due to suboptimal performance on atypical data
Solution Approach 1:
The system pre-calculates and stores similarity thresholds during the training phase, before actual inference occurs. This preliminary action allows the runtime system to make quick reliability decisions by simply comparing current data similarity against pre-computed thresholds, rather than performing complex analyses in real-time
Solution Approach 2:
A similarity measurement module is introduced as an intermediary component between the data input and ML model. This module computes similarity metrics and compares them against thresholds, acting as a mediator that determines whether atypical data should trigger model exit, thereby improving reliability without requiring the entire system to become complex
2Adaptability or versatility
If similarity threshold monitoring is implemented, then system adaptability improves for atypical data scenarios, but device complexity increases due to additional monitoring components
Solution Approach 1:
The system monitors and compares the similarity parameter between input data and training data against pre-determined thresholds. By changing and monitoring this key parameter, the system adapts to atypical data scenarios automatically, switching behavior based on whether the similarity parameter falls below threshold values
Solution Approach 2:
The system performs self-monitoring of data similarity and automatically determines when to exit or continue using the ML model based on pre-set thresholds. This self-service capability allows the system to adapt to atypical data without requiring external intervention or complex decision-making infrastructure
3Productivity
If ML model usage is extended without exit conditions, then productivity is maintained through continuous operation, but loss of information increases due to processing atypical data beyond model capabilities
Solution Approach 1:
The system implements feedback by continuously monitoring the similarity between input data and training data. When the similarity feedback indicates atypical data (below threshold), the system responds by exiting the ML model, preventing information loss from processing data outside the model's capabilities while maintaining productivity through appropriate model usage
Data Source
AI summary
Techniques and systems are provided for wireless communications. In some examples, a system receives input data for use by a first machine learning (ML) engine to generate an output result. The system may also receive information associated with a first set of training data for the first ML engine. The system may further determine a first dissimilarity amount between the received input data and the first set of training data based on the received information associated with the first set of training data. The system may also determine whether to continue using the first ML engine to generate the output result based on a comparison of the determined first dissimilarity amount to a first dissimilarity threshold.