DNN Layer and Filter Elimination for Compact Edge Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural network (DNN) model pruning techniques reduce parameters to zero values, maintaining memory size and hindering deployment on resource-constrained devices.
Innovation Solution
A method and system for DNN model size reduction by iteratively pruning, eliminating layers and filters based on sparsity statistics, guided by target size and performance constraints, to create a compact model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional pruning techniques are used to reduce parameters to zero values, then model performance is maintained, but model size in memory remains unchanged
Solution Approach 1:
The patent extracts and removes entire layers and filters from the neural network model based on sparsity statistics, rather than merely setting individual parameters to zero. This extraction approach actually reduces the model's memory footprint by eliminating redundant structural components while preserving essential functionality.
Solution Approach 2:
The patent changes the structural parameters of the model by dynamically adjusting the number of layers and filters based on calculated sparsity statistics. This parameter transformation converts a dense model into a sparse, compact model with reduced memory requirements while maintaining performance through intelligent structural modification.
2Volume of stationary object
If model size is reduced for edge device deployment, then memory constraints are satisfied, but model performance may deteriorate
Solution Approach 1:
The patent implements a feedback mechanism where sparsity statistics are calculated from the model's weight parameters, and this information feeds back into the model restructuring process. The sparsity metrics guide the elimination of layers and filters, ensuring that performance-critical components are preserved while redundant elements are removed, thus maintaining performance during size reduction.
Solution Approach 2:
The patent performs preliminary analysis of weight parameter distributions and calculates sparsity statistics before executing model reduction. This preliminary characterization of the model's structural properties enables informed decisions about which layers and filters to eliminate, preventing performance degradation by preserving essential model components.
3Volume of stationary object
If iterative model reduction is performed with layer and filter elimination, then model size is significantly reduced, but computational complexity of the process increases
Solution Approach 1:
The patent segments the model reduction process into distinct iterative steps: calculating sparsity statistics for individual layers, evaluating which layers exceed sparsity thresholds, and selectively eliminating those layers. This segmentation of the complex reduction process into manageable stages makes the computational complexity more tractable and systematic.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention generally relates to the field of deep learning, and, more particularly, to a method and system for deep neural network model size reduction by layer and filter elimination. Conventional methods result in zero weight parameters but does not effectively reduce the size of the models. Thus, the method of present disclosure imbibes capability of effective discovery of sparsified layers and systematically generates smaller size models close to the performance of base model. Further, the disclosed method gains regularization effect due to the removal of unnecessary parameters. The smaller size model reduces computational burden, energy consumption, and latency along with satisfying smaller memory requirements and thus, it leads to the possibility of deploying at different edge devices that provides the opportunity to develop niche and important applications in different domains including automobiles, robotics, healthcare.