Auxiliary Neural Network for Task Compatibility Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining task compatibility in shared neural networks are computationally expensive and time-consuming, making them impractical for real-life applications, especially in resource-constrained environments like vehicles, where efficient multitask learning is necessary for image processing tasks.
Innovation Solution
A method involving training separate neural networks for individual tasks, forming an estimation neural network with an auxiliary network to calculate an information share measure indicating overlap between task outputs, and using a threshold to determine suitable task clusters for joint processing in a neural network, thereby reducing computational effort while maintaining high performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all task combinations are explicitly trained to measure task compatibility, then measurement precision is improved, but loss of time and use of energy increase significantly
Solution Approach 1:
The patent trains individual neural networks for each task beforehand and stores their outputs. During compatibility measurement, these pre-trained networks are used to generate encoded information without requiring retraining, thus saving significant training time while maintaining measurement precision.
Solution Approach 2:
The patent introduces an auxiliary neural network as an intermediary that takes the outputs of pre-trained task-specific networks and computes compatibility measures. This auxiliary network acts as a mediator that eliminates the need to retrain all task combinations while providing accurate compatibility assessment.
2Measurement precision
If all task combinations are explicitly trained to measure task compatibility, then measurement precision is improved, but use of energy increases significantly
Solution Approach 1:
The patent trains individual neural networks for each task beforehand and stores their outputs. During compatibility measurement, these pre-trained networks are used to generate encoded information without requiring retraining, thus saving significant computational energy while maintaining measurement precision.
Solution Approach 2:
The patent introduces an auxiliary neural network as an intermediary that takes the outputs of pre-trained task-specific networks and computes compatibility measures. This auxiliary network acts as a mediator that eliminates the need to retrain all task combinations while providing accurate compatibility assessment, thereby reducing energy consumption.
3Device complexity
If separate models are trained for each task cluster to meet hardware constraints, then device complexity is reduced, but productivity decreases due to inability to identify optimal task groupings
Solution Approach 1:
The patent computes compatibility measures for all task pairs and uses this feedback information to optimally group tasks into clusters. This feedback-driven approach ensures that tasks with high compatibility are grouped together, maximizing computational efficiency while maintaining manageable model complexity suitable for hardware deployment.
Data Source
AI summary
The example embodiments relate to a computer-implemented method for determining clusters of tasks, the clusters at least partially including multiple tasks to be executed in a joint encoder portion of a neural network. The embodiments suggest estimating information share measures based on an auxiliary neural network in order to determine clusters of tasks to be executed in a joint encoder portion of a neural network.