Mobile Style Transfer DNN Layer Reduction for Real-Time Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing DNN models are complex and resource-intensive, making real-time inference in mobile environments challenging due to high computational demands, unnecessary layers, and resource mismatches between resource-abundant and resource-limited environments, leading to potential overflows.
Innovation Solution
Optimize DNN models for mobile environments by reducing the number of deep layers through transfer learning, decreasing feature sizes, and incorporating instance normalization and scaling in residual layers to prevent overflows.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If DNN models use many deep layers to improve accuracy, then model precision is improved, but device complexity and computational resource requirements increase
Solution Approach 1:
The patent segments the DNN model into essential and non-essential layers, retaining only the most critical deep layers for accuracy while removing redundant ones. This segmentation allows the model to maintain inference accuracy on mobile devices without the full complexity of the original architecture.
Solution Approach 2:
The patent extracts and removes unnecessary layers from the original DNN model, keeping only the essential components that contribute to accurate inference. This extraction process reduces model complexity and computational requirements while preserving the core functionality needed for mobile deployment.
2Productivity
If DNN models are optimized for resource-abundant environments, then model performance is improved, but adaptability to resource-limited mobile environments deteriorates
Solution Approach 1:
The patent changes key parameters of the DNN model including reducing the number of layers, adjusting feature map sizes, and modifying computational operations to be suitable for mobile environments. These parameter changes enable the model to adapt from resource-abundant to resource-limited environments while maintaining acceptable performance.
Solution Approach 2:
The patent creates a dynamic optimization process that adapts the model architecture based on the target deployment environment's resource constraints. The model structure is made flexible and adjustable rather than fixed, allowing it to be dynamically optimized for different mobile device capabilities.
3Measurement precision
If DNN models process large feature sizes to maintain accuracy, then inference precision is improved, but computational power requirements and energy consumption increase
Solution Approach 1:
The patent extracts and removes redundant feature processing operations from the model, keeping only the essential feature transformations needed for accurate inference. This reduces the computational load and energy consumption while maintaining the precision required for mobile applications.
Solution Approach 2:
The patent applies different feature processing strategies to different parts of the model based on their importance. Critical features maintain larger processing dimensions for accuracy, while less important features use reduced processing to save energy, creating a locally optimized feature processing architecture.
4Adaptability or versatility
If DNN models use complex architectures to improve functionality, then model capability is improved, but ease of operation and deployment in mobile environments deteriorates
Solution Approach 1:
The patent segments the complex model architecture into modular components, making it easier to deploy and operate on mobile devices. By dividing the model into essential functional blocks, the deployment process becomes more manageable and the model easier to operate within mobile system constraints.
Data Source
AI summary
A system includes at least one processor having a learning part for learning a deep neural network (DNN)-based style transfer model by using an image of a specific style to be learned, the style transfer model being a DNN model having an architecture in which the number of deep layers is reduced through transfer learning using a previously learned result.