Trained Model Validation Using Homogeneity Tests
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Models trained using supervised learning methods, such as neural networks, cannot be theoretically proven correct and are susceptible to misclassifications due to biased designs or attacks, posing a threat to efficiency and security in critical applications.
Innovation Solution
A method involving a computer system that compares the answers of multiple trained models to a test dataset, using homogeneity tests like probability distribution analysis or direct comparisons, to identify deviating models and take corrective actions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple models are trained using supervised learning methods, then the productivity and automation capability are improved, but the reliability and security deteriorate due to undetectable misclassifications and potential attacks
Solution Approach 1:
The patent applies local quality by treating each model's validation independently through individual homogeneity tests. Each model is tested against its own training data distribution to detect local deviations or misclassifications specific to that model, rather than treating all models uniformly. This allows targeted identification of problematic models while maintaining the overall system's productivity.
2Ease of operation
If statistical performance metrics are used to evaluate models, then the ease of operation is improved, but the measurement precision deteriorates because correctness cannot be theoretically proven
Solution Approach 1:
The patent implements feedback by using homogeneity tests that compare model predictions on training data against the expected training distribution. This creates a feedback loop where models are continuously validated against their own training characteristics, providing theoretical proof of correctness rather than relying solely on statistical performance metrics. The feedback mechanism detects deviations that indicate misclassifications or attacks.
3Reliability
If homogeneity tests are performed on multiple models, then the reliability and security are improved by detecting deviating models, but the device complexity and time consumption increase
Solution Approach 1:
The patent applies preliminary action by performing homogeneity tests during the model validation phase before deploying models to production. By conducting these tests in advance on training data, the system identifies and flags problematic models before they are used in critical applications, preventing security issues and misclassifications from affecting operational reliability.
Solution Approach 2:
The validation system is segmented into independent homogeneity tests for each model. Each model undergoes separate validation against its own training distribution, allowing the system to scale by adding or removing individual model validations without increasing overall system complexity. This modular approach handles multiple models efficiently.
4Productivity
If models are trained with biased designs or overtraining, then the productivity is improved through faster deployment, but the reliability deteriorates due to undetected misclassifications
Solution Approach 1:
The patent implements self-service by having each model validate itself against its own training data distribution through homogeneity tests. Models with biased designs or overtraining issues will show deviations in their self-validation, automatically flagging their own problems without requiring external manual review. This maintains rapid deployment while ensuring reliability through self-detection mechanisms.
Data Source
AI summary
A method for detecting a deviating model among a plurality of different models trained using a supervised learning method, said method being performed by a computer system programmed with the trained models and including acquiring a test dataset, presenting said test dataset to each of the trained models and generating answers of each trained model to said test dataset, performing at least one homogeneity test based on said answers generated by at least two models of the plurality of trained models, when said homogeneity test fails, performing a predetermined action indicative that one of said at least two models has been detected as deviating with regard to the other trained models.