Multi-Variant Inference Models With Staged Compute Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-variant inference models in information handling systems face high computation complexity and power inefficiency, leading to significant energy consumption and environmental impact.
Innovation Solution
Implementing an inference engine with optimized multiple inference stages and a modeler that refines these stages based on input evaluation, incorporating data cleansing, dimensional reduction, and compute reduction techniques to simplify data and create sub-cluster groupings, reducing the number of calculations required.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-variant inference models are used to improve accuracy, then prediction quality is improved, but computation complexity increases
Solution Approach 1:
The patent segments the inference model into multiple stages with varying levels of complexity. Different inference models are applied at different stages, with simpler models processing easier cases and more complex models handling difficult cases. This segmentation allows the system to achieve high prediction accuracy while avoiding unnecessary computational complexity for straightforward predictions.
Solution Approach 2:
The patent implements dynamic model selection where the system adaptively chooses which inference model to apply based on the specific input characteristics. The complexity of the inference model is dynamically adjusted according to the difficulty of the prediction task, allowing the system to maintain high accuracy while optimizing computational resources for each specific case.
2Productivity
If multi-variant inference models are deployed to improve performance, then prediction capability is improved, but power consumption increases
Solution Approach 1:
The patent segments the inference workload across multiple stages with different power requirements. By distributing complex inference tasks across multiple simpler stages rather than using a single complex model, the system achieves improved prediction capability while reducing peak power consumption and enabling better thermal management.
Solution Approach 2:
The patent employs periodic or staged inference actions where simpler models are applied first, and only when necessary does the system proceed to more power-intensive inference stages. This periodic approach to model execution allows the system to maintain high prediction capability while managing power consumption in a controlled, intermittent manner rather than continuous high-power operation.
Data Source
AI summary
An information handling system may define a first grouping of inputs to a first inference model, determine a first number of inference stages for the first inference model, and calculate an accuracy of an output of the first inference model. When the accuracy is within a threshold accuracy, the system may load the first inference model to multiple computing devices.