Variable Power Mode Machine Learning Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models require significant compute resources and power, limiting their deployment to mobile devices and low-power edge processing systems due to high power demands, which hampers energy efficiency and performance.
Innovation Solution
Implementing variable power mode machine learning models that selectively use lower power processing modes for reliable inferences and higher power modes when necessary, allowing the same model to operate in multiple power and performance modes without changing the underlying architecture, thereby improving both performance and energy efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning models are deployed to achieve optimal performance, then processing accuracy is improved, but power consumption increases
Solution Approach 1:
The patent implements dynamic power mode selection where the system transitions between different power modes (first power mode and second power mode) based on the characteristics of the input data. The machine learning model is configured to operate in a first power mode for standard processing and switches to a second power mode when specific conditions are met, such as when input data falls within a certain probability range. This dynamic adaptation allows the system to optimize the balance between processing accuracy and power consumption in real-time.
Solution Approach 2:
The system changes operational parameters by adjusting the power mode of the machine learning model based on input data characteristics. When the input data probability falls within a predetermined range (indicating higher uncertainty or complexity), the system switches to a second power mode with higher computational resources. For standard cases outside this range, the first power mode is used. This parameter change approach enables adaptive resource allocation that matches computational needs with actual data requirements.
2Reliability
If machine learning models operate at high power levels to ensure reliable inferences, then inference reliability is improved, but energy efficiency deteriorates
Solution Approach 1:
The patent applies partial action by using high-power processing only when necessary. Instead of always operating at full power to ensure reliability, the system selectively activates the second power mode only when input data probability indicates potential inference uncertainty. For the majority of standard cases, the lower-power first mode suffices, providing adequate reliability while conserving energy. This partial application of high-power processing optimizes the reliability-energy efficiency tradeoff.
3Use of energy by moving object
If variable power mode processing is implemented to improve energy efficiency, then power consumption is reduced, but system complexity increases
Solution Approach 1:
The patent implements a universal machine learning model that can operate in multiple power modes without requiring separate models for each mode. The same model architecture and parameters are used across both first and second power modes, with the only difference being the computational resources allocated. This multi-functionality approach reduces system complexity compared to maintaining separate models for different power levels, as the model itself remains unchanged while adapting to different operational conditions.
Data Source
AI summary
Certain aspects of the present disclosure provide a method for performing multimode inferencing, comprising: receiving machine learning model input data from a requestor; processing the machine learning model input data with a machine learning model using processing hardware at a first power level to generate first output data; selecting a second power level for the processing hardware based on comparing the first output data to a threshold value; processing the machine learning model input data with the machine learning model using the processing hardware at the second power level to generate second output data; and sending second output data to the requestor.