Neural Network Channel Compression via Graph-Based Redundancy Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks require significant computational resources and memory bandwidth due to large sets of coefficients, particularly in fully-connected and convolution layers, leading to high processing demands and memory bandwidth requirements.
Innovation Solution
A method and system for compressing neural networks by identifying and removing redundant channels through a graph traversal process, forming a graph that represents data flow, and eliminating channels that do not affect the output, thereby reducing the number of coefficients and computational demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural networks use large sets of coefficients in fully-connected and convolution layers, then computational accuracy and model capacity are improved, but computational load and memory bandwidth requirements increase
Solution Approach 1:
The patent extracts and removes redundant channels from the neural network by identifying channels that do not contribute to the output through graph traversal analysis. This extraction principle directly addresses the contradiction by eliminating unnecessary computational operations while preserving the essential functionality and accuracy of the network.
Solution Approach 2:
The patent changes the parameter of channel dimensionality by reducing the number of active channels in fully-connected and convolution layers. Through graph-based analysis of data flow dependencies, the system modifies the network architecture parameters to remove redundant channels, thereby reducing computational load while maintaining accuracy.
2Adaptability or versatility
If neural networks use large sets of coefficients, then model capacity and representational power are improved, but memory bandwidth requirements increase
Solution Approach 1:
The patent extracts redundant channels that do not contribute to the output through graph traversal analysis. By removing these unnecessary channels, the system reduces the quantity of data that must be transferred through memory bandwidth while preserving the essential model capacity and representational power needed for accurate inference.
Solution Approach 2:
The patent segments the neural network into essential channels and redundant channels by analyzing data flow dependencies through graph traversal. This segmentation allows the system to retain only the necessary channels for maintaining model capacity while discarding redundant ones that consume memory bandwidth.
3Productivity
If redundant channels are removed from the neural network, then computational load and memory bandwidth requirements are reduced, but network functionality may be compromised
Solution Approach 1:
The patent employs graph traversal analysis to feedback-check the impact of removing channels on the network output. By analyzing data flow dependencies and tracking how channels propagate information through the network, the system can confidently remove only truly redundant channels while preserving functionality through continuous verification of data flow paths.
Solution Approach 2:
The patent uses graph traversal analysis as an intermediary mechanism to determine channel redundancy. This intermediary analysis layer examines data flow dependencies and provides informed decisions about which channels can be safely removed, ensuring that network functionality is preserved while achieving computational efficiency.
Data Source
Figure 1~2a
Figure 2b
Figure 2c
AI summary
A computer implemented method of compressing a neural network, the method comprising: receiving a neural network comprising a plurality of layers; forming a graph that represents the flow of data through the plurality layers of the neural network, the graph comprising: a plurality of vertices, each vertex of the plurality of vertices being representative of an output channel of a layer of the plurality of layers of the neural network; and one or more edges, each edge of the one or more edges representing the potential flow of non-zero data between respective output channels represented by a respective pair of vertices; identifying, by traversing the graph, one or more redundant channels comprised by the plurality of layers of the neural network; and outputting a compressed neural network in which the identified one or more redundant channels are not present in the compressed neural network.