Rotating Neural Network Subsets for Perception
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Ensemble neural networks used for perception tasks in electronic devices and vehicles face challenges with increased computational time and resource requirements, making them unsuitable for real-time applications and resource-limited hardware.
Innovation Solution
A rotating scheme is implemented where only a subset of neural networks is executed at each time instance, with aggregated outputs from different subsets of neural networks combined over consecutive time instances to maintain performance while reducing computational time and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ensemble neural networks are used to improve accuracy and robustness, then perception performance is improved, but computational time and processing power requirements increase significantly
Solution Approach 1:
The ensemble of neural networks is segmented into multiple subsets that are executed at different time instances. Instead of running all ensemble networks simultaneously on every input, the system divides the ensemble into N subsets and rotates through them across N consecutive time instances, processing only one subset per time instance while maintaining the benefits of ensemble diversity.
Solution Approach 2:
The system implements periodic action by rotating through different subsets of the neural network ensemble in a cyclic manner across consecutive time instances. Each subset is executed periodically, ensuring that all ensemble networks contribute to the overall perception task while distributing the computational load over time, thus maintaining accuracy without requiring all networks to run simultaneously.
2Reliability
If ensemble neural networks are deployed, then accuracy and robustness are improved, but hardware resource requirements increase
Solution Approach 1:
The ensemble network is segmented into multiple subsets that can be executed sequentially rather than simultaneously. This segmentation allows the system to maintain the robustness benefits of using multiple diverse neural networks while reducing the peak hardware resource requirements, as only one subset needs to be active at any given time instance.
Solution Approach 2:
The system dynamically switches between different subsets of the neural network ensemble across time instances. This dynamic rotation allows the hardware resources to be reused for different network subsets at different times, effectively reducing the total hardware footprint required compared to having all ensemble networks active simultaneously, while still achieving the robustness benefits of ensemble diversity.
Data Source
AI summary
The present invention relates to a method for performing a perception task of an electronic device or a vehicle, using a plurality of neural networks trained to generate a perception output based on an input image, wherein at least two neural networks of the plurality of neural networks are different from each other. The method includes: for a time instance of a plurality of consecutive time instances: obtaining an image depicting a portion of a surrounding environment of the electronic device or the vehicle; processing the image associated with the time instance using a subset of neural network(s) to obtain a network output for the time instance; and determining an aggregated network output by combining the obtained network output for the time instance with network outputs obtained for a number of preceding time instances.