Neural Network Model Simplification via Layer Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As neural network models increase in complexity and size, the inference time and power consumption also increase, necessitating a method to simplify these models without hardware upgrades to maintain performance.
Innovation Solution
A simplification method that reduces multiple linear operation layers to at most two layers by calculating new weights and biases from the original model, allowing for equivalent inference performance with reduced computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of linear operation layers in the neural network model is increased to improve model accuracy, then the model accuracy is improved, but the inference time increases
Solution Approach 1:
The patent merges multiple linear operation layers into a single equivalent linear operation layer by calculating new weight and bias parameters that preserve the functional behavior of the original multi-layer structure. This consolidation reduces the number of sequential operations during inference while maintaining model accuracy.
Solution Approach 2:
The patent performs preliminary calculations during the simplification phase to pre-compute the equivalent weight and bias parameters for the consolidated layer. This preliminary action eliminates the need for multiple sequential linear operations during inference, thereby reducing inference time without sacrificing accuracy.
2Measurement precision
If the number of linear operation layers in the neural network model is increased to improve model accuracy, then the model accuracy is improved, but the power consumption increases
Solution Approach 1:
The patent merges multiple linear operation layers into a single equivalent linear operation layer by calculating new weight and bias parameters that preserve the functional behavior of the original multi-layer structure. This consolidation reduces the number of sequential operations during inference while maintaining model accuracy.
Solution Approach 2:
The patent performs preliminary calculations during the simplification phase to pre-compute the equivalent weight and bias parameters for the consolidated layer. This preliminary action eliminates the need for multiple sequential linear operations during inference, thereby reducing inference time without sacrificing accuracy.
3Adaptability or versatility
If the number of linear operation layers in the neural network model is increased to improve model complexity, then the model capacity is improved, but the device complexity increases
Solution Approach 1:
The patent merges multiple linear operation layers into a single equivalent linear operation layer by calculating new weight and bias parameters that preserve the functional behavior of the original multi-layer structure. This consolidation reduces the number of sequential operations during inference while maintaining model accuracy.
Solution Approach 2:
The patent transforms the structural complexity of multiple layers into parameter complexity by computing equivalent weight and bias parameters. The simplified model has fewer layers but uses pre-calculated parameters that encapsulate the transformation logic, effectively trading structural complexity for parameter complexity.
Data Source
AI summary
A simplification device and a simplification method for neural network model are provided. The simplification method may simplify an original trained neural network model to a simplified trained neural network model, wherein the simplified trained neural network model includes at most two linear operation layers. The simplification method includes: converting the original trained neural network model into an original mathematical function; performing an iterative analysis operation on the original mathematical function to simplify the original mathematical function to a simplified mathematical function, wherein the simplified mathematical function has a new weight; computing the new weight by using multiple original weights of the original trained neural network model; and converting the simplified mathematical function to the simplified trained neural network model.