Policy Shims for ML/AI Controllability and Modifiability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is a lack of visibility, flexibility, and controllability in machine learning (ML) and artificial intelligence (AI) applications, leading to resistance and uncertainty among operators and users due to the 'black box' nature of these systems, making it difficult to modify or debug ML/AI algorithms once deployed.
Innovation Solution
The introduction of policy shims that can be applied to ML/AI modules to provide input and output controls, allowing for filtering, modification, and augmentation of data streams, enabling explicit control and flexibility through a centralized or distributed policy decision engine.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If ML/AI algorithms are used to automate decision-making, then productivity and automation capability are improved, but controllability and user confidence deteriorate due to the black box nature of these systems
Solution Approach 1:
The patent introduces policy shims as intermediary components that sit between users and ML/AI algorithms. These shims provide a simplified interface layer that translates high-level policy intentions into algorithm-specific parameters, enabling users to control ML/AI systems without needing to understand their internal workings. This resolves the contradiction by maintaining automation capability while improving controllability through the mediator layer.
Solution Approach 2:
The patent segments the control mechanism into modular policy shims that can be independently configured and applied to different ML/AI algorithms. Each shim represents a discrete policy component that can be adjusted without affecting the entire system. This segmentation enables fine-grained control over automated systems, resolving the contradiction between maintaining automation and improving ease of operation.
2Productivity
If ML/AI algorithms are deployed for complex tasks, then productivity is improved, but the ability to modify and debug algorithms deteriorates due to their complexity and training time
Solution Approach 1:
The patent extracts the policy control logic from the ML/AI algorithms themselves and places it in separate policy shim components. This extraction allows policy parameters to be modified independently of the algorithm's core functionality and training. Users can adjust policy shims without retraining algorithms, resolving the contradiction between maintaining task execution capability and improving modifiability.
Solution Approach 2:
The patent establishes policy shims as pre-configured control layers that are set up before algorithm execution. These shims contain pre-defined policy rules and parameters that can be adjusted without triggering algorithm retraining. This preliminary action structure enables easy modification of algorithm behavior while maintaining the deployed algorithm's task execution capabilities.
3Ease of operation
If policy controls are added to ML/AI systems to improve controllability, then ease of operation is improved, but device complexity increases due to additional control layers
Solution Approach 1:
The patent designs policy shims as universal components that can be applied to multiple different ML/AI algorithms through a common interface. Rather than creating separate control mechanisms for each algorithm, the same shim framework serves multiple algorithms, reducing overall system complexity while maintaining improved controllability. This universality resolves the contradiction by providing enhanced ease of operation without proportionally increasing device complexity.
Data Source
AI summary
A method includes creating one or more first policy shims to be applied to a ML/AI module, applying the one or more first policy shims to an input or an output of the ML/AI module and executing the ML/AI module on a data set in response to the applying step. The one or more first policy shims includes an input policy shim and an output policy shim and the applying step includes applying the input policy shim to the data set prior to the executing step and applying the output policy shim to an output of the executing step.