Vehicle Neural Network Resource Allocation Under Thermal Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for managing computing resources in vehicle systems executing multiple neural networks often result in across-the-board reductions, compromising safety-critical functions by treating all neural networks equally, despite varying importance for safe vehicle operations.
Innovation Solution
A method that dynamically allocates computing resources to neural networks based on their relative importance to overall vehicle safety performance, adjusting hyperparameters according to a performance-effectiveness curve and reallocating resources based on dynamic availability and usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Temperature
If computing resources are reduced to manage thermal constraints, then temperature is controlled, but safety-critical neural network performance deteriorates
Solution Approach 1:
The patent applies local quality by differentiating resource allocation across different neural networks based on their safety criticality. Instead of uniformly reducing resources, the system identifies and protects safety-critical neural networks while allowing non-critical ones to receive reduced resources, thereby controlling temperature while maintaining essential safety functions.
Solution Approach 2:
The patent segments neural networks into safety-critical and non-safety-critical categories. This segmentation enables selective resource management where thermal constraints are enforced on non-critical networks while safety-critical networks maintain adequate resources, resolving the contradiction between temperature control and safety performance.
2Use of energy by moving object
If computing resources are reduced to manage power consumption, then energy efficiency improves, but overall neural network accuracy deteriorates
Solution Approach 1:
The patent applies local quality by allocating power resources differentially to neural networks based on their safety criticality. Safety-critical networks receive sufficient power to maintain high accuracy, while non-critical networks undergo power reduction, thereby improving overall energy efficiency without compromising essential safety functions.
Solution Approach 2:
The patent segments the neural network workload into safety-critical and non-safety-critical groups, enabling selective power management. This segmentation allows the system to reduce power consumption by throttling non-critical networks while preserving accuracy in safety-critical networks, resolving the contradiction between energy efficiency and inference accuracy.
3Productivity
If computing resources are reduced to manage resource constraints, then system load decreases, but vehicle safety performance deteriorates
Solution Approach 1:
The patent applies local quality by selectively managing resources for different neural networks based on their contribution to safety performance. Safety-critical networks are identified and protected from resource reduction, while non-critical networks receive reduced resources to alleviate overall system load, thereby maintaining safety performance while improving productivity.
Solution Approach 2:
The patent segments neural networks by their safety criticality level, enabling differential resource allocation. This segmentation allows the system to reduce load on non-critical networks while ensuring safety-critical networks maintain adequate resources, resolving the contradiction between productivity and safety performance.
Data Source
AI summary
Embodiments include methods performed by a processor of a vehicle for allocating processing resources to concurrently-executing neural networks. The methods may include determining a priority of each of a plurality of neural networks executing on a vehicle processing system based on a contribution of each neural network to overall vehicle safety performance, and allocating computing resources to the plurality of neural networks based on the determined priority of each neural network. In some embodiments, the methods may dynamically adjust hyperparameters of one or more neural networks.