Neural Network Layer Reconfiguration for Hardware Utilization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep neural networks (DNNs) lack an optimal structure that efficiently processes tasks without considering hardware characteristics, leading to suboptimal performance due to inefficient resource utilization.
Innovation Solution
A processor-implemented method that calculates required resource information for each neural network layer, determines reconfiguration target layers based on hardware information, and selects and implements reconfigured layers to optimize performance by adjusting the number of paths and kernel channels according to processing core and memory capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If DNN models are designed without considering hardware characteristics, then model flexibility and task adaptability are maintained, but hardware utilization efficiency and processing performance deteriorate
Solution Approach 1:
The patent applies dynamics by making the neural network model structure adjustable and reconfigurable based on hardware characteristics. The system dynamically determines reconfiguration layers, reconfiguration paths, and kernel channel numbers according to different hardware platforms (processing core information, memory information), allowing the model to adapt its structure to match hardware capabilities while maintaining task versatility
Solution Approach 2:
The patent changes key structural parameters of the neural network model including the number of reconfiguration layers, reconfiguration paths per layer, and kernel channel numbers. These parameters are adjusted based on hardware characteristics (number of processing cores, memory size, bandwidth) to optimize the balance between model flexibility and hardware utilization efficiency
2Productivity
If the number of reconfiguration paths is increased to improve hardware utilization, then processing throughput increases, but computational complexity and latency may worsen
Solution Approach 1:
The patent applies partial action by selectively reconfiguring only certain layers (reconfiguration target layers) of the neural network model rather than all layers. The system determines which layers to reconfigure based on hardware characteristics and task requirements, applying reconfiguration paths to specific layers where it provides the most benefit while avoiding unnecessary complexity in other layers
Solution Approach 2:
The system dynamically adjusts the number of reconfiguration paths based on hardware capabilities. The reconfiguration path determination module calculates the optimal number of paths according to processing core information and memory characteristics, allowing the system to maximize throughput while maintaining acceptable latency by adapting to available computational resources
3Measurement precision
If kernel channel numbers are increased to enhance model accuracy, then classification performance improves, but memory requirements and computational load increase
Solution Approach 1:
The patent changes the kernel channel number parameter based on memory characteristics and hardware capabilities. The reconfiguration layer determination module adjusts kernel channel numbers to optimize the balance between classification accuracy and memory requirements, ensuring that the model achieves sufficient precision without exceeding available memory resources
Solution Approach 2:
The patent applies local quality by differentiating kernel channel configurations across different layers and reconfiguration paths. Instead of using uniform kernel channels throughout the model, the system adjusts kernel channel numbers locally in specific reconfiguration layers based on their specific computational requirements and hardware constraints, optimizing both accuracy and memory usage
Data Source
AI summary
A processor-implemented neural network model reconfiguration method is provided. The method calculates required resource information of each of a plurality of layers of a neural network model, determines whether a layer of the plurality of layers is a reconfiguration target layer based on the required resource information and hardware information, determines reconfiguration layers, with respect to the reconfiguration target layer, based on required resource information of the reconfiguration target layer and the hardware information, selects one of the reconfiguration target layer and the determined reconfiguration layers based on performance information of the reconfiguration target layer and performance information of each of the determined reconfiguration layers, and implements one of the determined reconfigured layers in the neural network model based on the selecting of the determined reconfiguration layers.