Fully Connected Layer to Convolution Transformation for Edge Inference Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models like the Transformer experience processing speed decreases due to frequent dimensional transformations and inverse transformations, leading to slower performance compared to convolutional neural networks (CNNs), especially in edge devices like smartphones.
Innovation Solution
An information processing device that acquires a deep learning model, selects a fully connected layer, transforms it into a convolution layer, and deletes dimensional transformation and inverse transformation layers, resulting in a second model with fewer layers that can process information more quickly without compromising accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If dimensional transformation layers and inverse transformation layers are included in the deep learning model to enable fully connected layers to process two-dimensional information, then the model can maintain accuracy, but the processing speed decreases due to frequent dimensional transformations
Solution Approach 1:
The patent extracts and removes the dimensional transformation layer and dimensional inverse transformation layer from the deep learning model. By taking out these specific components that cause frequent dimensional transformations, the model eliminates the processing speed bottleneck while preserving the core functionality of the fully connected layer through alternative implementation
Solution Approach 2:
The patent changes the operational dimension of the fully connected layer from two-dimensional to three-dimensional processing. Instead of transforming dimensions back and forth, the invention processes information directly in three-dimensional space using the fully connected layer, eliminating the need for dimensional transformation layers while maintaining or improving processing efficiency
2Adaptability or versatility
If multiple layers including dimensional transformation and inverse transformation layers are used in the deep learning model, then the model architecture can handle dimensionality changes, but the number of layers increases leading to slower processing
Solution Approach 1:
The patent extracts and removes the dimensional transformation layer and dimensional inverse transformation layer from the deep learning model. By taking out these specific components that cause frequent dimensional transformations, the model eliminates the processing speed bottleneck while preserving the core functionality of the fully connected layer through alternative implementation
Solution Approach 2:
The patent merges the dimensionality handling capability directly into the fully connected layer by enabling it to process three-dimensional information natively. This consolidation eliminates the need for separate transformation layers, reducing the total number of layers while maintaining the model's ability to handle dimensional changes
Data Source
AI summary
An information processing device includes an input acquirer which acquires a first model of deep learning and a transform unit which selects a fully connected layer included in the first model, transforms the selected fully connected layer into a convolution layer, and deletes a dimensional transformation layer and a dimensional inverse transformation layer that are included in the first model. The dimensional transformation layer transforms the total number of dimensions of input information from three to two, and outputs the input information represented in two dimensions to the fully connected layer. The dimensional inverse transformation layer inversely transforms the number of dimensions of output information output from the fully connected layer from two to three.