Multi-Model ML Architecture for Reliable AI Under Compute Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous systems face challenges in achieving high algorithmic performance and dependability due to limited computing resources and super-exponential growth in AI compute requirements, leading to ML model failures and safety concerns.
Innovation Solution
A multi-model Machine Learning (ML) architecture is constructed using an efficient Neural Architecture Search (NAS) technique, incorporating diverse ensemble members and stochastic models, to enhance algorithmic performance and dependability by leveraging edge/cloud servers for compute-intensive tasks while optimizing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single ML model is used to manage visual perception tasks, then the system complexity is low, but the algorithmic performance and dependability are insufficient
Solution Approach 1:
The patent divides the ML system into multiple specialized models (e.g., primary model, secondary model, fail-safe model) that each handle specific aspects of visual perception tasks. This segmentation allows the system to achieve higher reliability through model diversity and redundancy while managing complexity through modular architecture and specialized function assignment.
Solution Approach 2:
The patent implements fail-safe and secondary models as pre-prepared backup systems that activate when the primary model fails or produces unreliable results. This beforehand cushioning ensures dependability by having predetermined fallback mechanisms in place before failures occur, addressing the reliability requirement without requiring complex real-time decision-making.
2Reliability
If multiple ML models are deployed to improve performance and safety, then reliability increases, but computing resources are exceeded
Solution Approach 1:
The patent implements a hierarchical model execution strategy where the system uses the lightweight primary model for most operations (partial action), and only activates heavier secondary or fail-safe models when specifically needed for safety-critical decisions or when the primary model's confidence is insufficient. This approach achieves the necessary reliability without consistently consuming excessive computing resources.
Solution Approach 2:
The patent assigns different computational characteristics to different models based on their specific functions: the primary model is optimized for speed and low resource usage for routine tasks, while secondary and fail-safe models are optimized for accuracy and safety but activated only when needed. This local quality optimization allows the system to meet reliability requirements without uniformly high resource consumption across all operations.
3Reliability
If model complexity is increased to improve accuracy, then algorithmic performance improves, but resource constraints are violated
Solution Approach 1:
The patent implements dynamic model selection where the system adapts which model to use based on real-time conditions such as task complexity, environmental factors, and confidence levels. This dynamics allows the system to use simpler models when they suffice (maintaining resource efficiency) and switch to more complex models only when accuracy requirements demand it, thus resolving the contradiction between accuracy and resource efficiency.
Data Source
AI summary
For example, an apparatus may include an input to receive Machine Learning (ML) model information corresponding to an ML model to process input information; and a processor to construct a multi-model ML architecture including a plurality of ML model variants based on the ML model, wherein the processor is configured to determine the plurality of ML model variants based on an attribution-based diversity metric corresponding to a model group including a first ML model variant and a second ML model variant, wherein the attribution-based diversity metric corresponding to the model group is based on a diversity between a first attribution scheme and a second attribution scheme, the first attribution scheme representing first portions of the input information attributing to an output of the first ML model variant, the second attribution scheme representing second portions of the input information attributing to an output of the second ML model variant.