Deep Neural Network Layer and Filter Elimination for True Size Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural network (DNN) model pruning techniques reduce parameters to zero but do not effectively decrease model size, hindering deployment on resource-constrained devices due to memory occupancy by zero-valued weights.
Innovation Solution
A method and system for DNN model size reduction by iteratively pruning and eliminating layers and filters based on sparsity statistics, guided by performance and size constraints, to create a compact model suitable for edge devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional pruning techniques are used to set weights to zero, then model accuracy is maintained, but model size cannot be reduced due to memory occupancy by zero-valued weights
Solution Approach 1:
The patent extracts and removes entire layers and filters from the neural network model based on sparsity statistics. By identifying and eliminating layers with insufficient sparsity or critical functionality, the model achieves actual size reduction while maintaining accuracy, rather than merely setting weights to zero which occupies memory.
Solution Approach 2:
The patent changes the structural parameters of the model by dynamically adjusting the number of layers and filters based on sparsity statistics. This involves modifying the model architecture itself rather than just adjusting weight values, enabling true model size reduction while preserving performance.
2Reliability
If large neural network models are used for time series classification, then model performance is improved, but deployment on edge devices is hindered due to memory constraints
Solution Approach 1:
The patent segments the neural network model into individual layers and filters, allowing selective elimination of less important components. This segmentation enables the model to be reduced to a compact version suitable for edge devices while maintaining core functionality for time series classification.
Solution Approach 2:
The patent extracts and removes unnecessary layers and filters from the model, creating a streamlined version optimized for deployment on resource-constrained edge devices. This extraction process maintains essential model capabilities while reducing memory requirements.
3Quantity of substance
If iterative pruning and layer elimination are performed, then model size is reduced, but computational complexity increases during the optimization process
Solution Approach 1:
The patent employs self-service mechanisms where the model automatically evaluates its own sparsity statistics and identifies which layers or filters can be eliminated. This self-evaluation process reduces the need for external intervention and manual optimization, simplifying the overall process despite the iterative nature of model reduction.
Data Source
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.