Integrated Sensor and Hyperparameter Optimization for ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In machine learning environments utilizing sensor data, optimizing only hyperparameters limits learning efficiency, and there is a need to integrate optimization of sensor parameters and preprocessing filters to enhance performance.
Innovation Solution
A method that integrates the optimization of sensor parameters, hyperparameters, and preprocessing filters by performing training on a machine learning model until a predetermined termination condition is satisfied, selecting and determining optimized parameters and filters that minimize the loss value, using a system comprising an optimization engine, sensor manager, data filter manager, machine learning manager, and optimization manager.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If only hyperparameters of machine learning model are optimized, then optimization process is simple, but learning efficiency enhancement is limited
Solution Approach 1:
The patent merges sensor parameter optimization, preprocessing filter optimization, and hyperparameter optimization into a single integrated optimization process. The optimization manager simultaneously adjusts sensor parameters (e.g., sampling frequency, resolution), preprocessing filter parameters (e.g., filter coefficients, window sizes), and machine learning model hyperparameters (e.g., learning rate, batch size) to collectively minimize the loss function, rather than optimizing them separately as in conventional approaches.
Solution Approach 2:
The optimization manager performs multiple optimization functions simultaneously: it optimizes sensor parameters for data acquisition, preprocessing filter parameters for data cleaning and feature extraction, and hyperparameters for model training. This multi-functional optimization approach enables a single system to handle all parameter tuning needs, improving learning efficiency without proportionally increasing system complexity.
2Productivity
If sensor parameters and preprocessing filters are integrated into optimization, then learning performance is improved, but optimization system complexity increases
Solution Approach 1:
The optimization manager serves as an intermediary component that coordinates between sensor management, data filtering, and machine learning model training. It receives feedback from the loss function and adjusts parameters across all three domains (sensor, preprocessing, and model) in a unified manner, simplifying the overall optimization system architecture while achieving improved learning performance.
Solution Approach 2:
The system implements a feedback mechanism where the loss function output from machine learning model training is fed back to the optimization manager. The optimization manager uses this feedback to iteratively adjust sensor parameters, preprocessing filter parameters, and hyperparameters, creating a closed-loop optimization system that automatically converges to optimal parameter settings without requiring complex manual tuning.
3Adaptability or versatility
If multiple types of sensors are utilized, then sensing capability is enhanced, but sensor parameter optimization becomes necessary
Solution Approach 1:
The optimization manager merges the optimization of parameters from multiple different sensor types (e.g., LIDAR, cameras, radar) into a unified optimization process. It simultaneously adjusts parameters across heterogeneous sensors along with preprocessing and model parameters, treating them as an integrated system rather than separate optimization problems, thereby managing complexity while maintaining enhanced sensing capability.
Data Source
AI summary
Disclosed is a method of integrally optimizing different types of parameters that require setting during a machine learning process. The disclosed method of integrally optimizing the parameters includes performing training on a machine learning model by selecting sensor parameters and machine learning model hyperparameters until a predetermined termination condition is satisfied; and determining, among the selected sensor parameters and machine learning model hyperparameters, an optimized sensor parameter and optimized machine learning model hyperparameter that minimize a loss value for the machine learning model, wherein the performing of the training on the machine learning model includes selecting the sensor parameters and machine learning model hyperparameters that satisfy a predetermined optimization range, and performing training on the machine learning model based on sensor data provided from a sensor by the selected sensor parameters and the machine learning model hyperparameters.