Convolutional Neural Networks with Runtime Adjustable Feature Resolutions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning models, such as neural networks, face challenges in meeting dynamic computational complexity requirements due to variability in available computing resources across different devices or over time, leading to inefficiencies in resource consumption and accuracy.
Innovation Solution
The implementation of a convolutional neural network with adjustable feature resolutions at runtime, where down-sampling layers are reconfigured within the network to optimize computational complexity based on available resources, allowing for dynamic adjustment of feature map resolutions and filter dilation levels to meet varying computational demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional fixed architecture neural networks are used, then model accuracy can be maintained, but computational complexity cannot be adapted to varying resource availability
Solution Approach 1:
The patent implements dynamic architecture by making the network structure adjustable at runtime through configurable down-sampling layer positions and feature map resolutions. The system dynamically reconfigures the computational graph based on available resources, transforming a static neural network into an adaptive system that can modify its own architecture during execution.
Solution Approach 2:
The patent changes key structural parameters of the neural network including down-sampling layer positions, feature map resolutions, and filter configurations. By adjusting these parameters based on computational resource availability, the system achieves adaptability without requiring complete model redesign, allowing runtime optimization of the architectural configuration.
2Measurement precision
If high resolution feature maps are processed throughout the network, then accuracy is improved, but computational complexity increases significantly
Solution Approach 1:
The patent segments the neural network into distinct stages with different feature map resolutions. By dividing the network architecture into segments with varying resolution requirements and strategically placing down-sampling layers, the system processes high-resolution features only where necessary while using lower resolutions elsewhere, reducing overall computational complexity while maintaining accuracy.
Solution Approach 2:
The patent applies local quality by allowing different parts of the network to operate at different feature map resolutions based on their specific functional requirements. Critical regions maintain high resolution for accuracy while non-critical regions use lower resolutions to reduce computation, creating a heterogeneous resolution architecture optimized for both precision and efficiency.
3Power
If down-sampling layers are placed early in the network, then computational complexity is reduced, but feature resolution and accuracy may be compromised
Solution Approach 1:
The patent dynamically determines the optimal position of down-sampling layers based on the specific computational constraints and accuracy requirements of each application scenario. Rather than using a fixed placement strategy, the system adapts the down-sampling positions at runtime to achieve the best balance between reducing computational complexity and preserving necessary feature resolution.
Data Source
AI summary
In some embodiments, an application receives a request to execute a convolutional neural network model. The application determines the computational complexity requirement for the neural network based on the computing resource available on the device. The application further determines the architecture of the convolutional neural network model by determining the locations of down-sampling layers within the convolutional neural network model based on the computational complexity requirement. The application reconfigures the architecture of the convolutional neural network model by moving the down-sampling layers to the determined locations and executes the convolutional neural network model to generate output results.