Pruned Neural Network Filters for Edge Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object recognition systems using deep learning models like convolutional neural networks (CNNs) require significant computing resources, making them unsuitable for resource-constrained devices such as consumer electronics, which compromises performance and accuracy.
Innovation Solution
The development of pruned neural networks customized for each edge node by a centralized server, where filters responsive to specific objects in reference images are selected, allowing for efficient object recognition with minimal resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models (CNNs) are used for object recognition, then recognition accuracy is improved, but computing resource requirements increase
Solution Approach 1:
The patent extracts and removes unnecessary filters from the parent neural network to create a pruned neural network. By analyzing which filters are actually needed for specific object recognition tasks and removing the rest, the system maintains high accuracy while significantly reducing computing resource consumption on edge nodes
Solution Approach 2:
The patent applies local quality by customizing the pruned neural network for each edge node based on its specific reference images and object recognition needs. Each edge node receives a tailored model with filters optimized for its local context, rather than using a generic full model, thereby reducing computational requirements while maintaining accuracy
2Adaptability or versatility
If full parent neural network is deployed to edge nodes, then comprehensive object detection capability is improved, but device complexity increases
Solution Approach 1:
The patent extracts only the necessary filters from the parent neural network and deploys them to edge nodes. This extraction process creates a simplified pruned neural network that maintains the ability to detect relevant objects while reducing model complexity and size for deployment on resource-constrained devices
Solution Approach 2:
The patent segments the parent neural network into individual filters and selectively chooses which ones to deploy. By dividing the comprehensive model into discrete filter components and only deploying necessary ones, the system reduces device complexity while preserving essential detection capabilities
3Productivity
If pruned neural network is used on edge nodes, then inference speed is improved, but accuracy may be reduced
Solution Approach 1:
The patent applies local quality by customizing the pruned neural network for each edge node based on its specific reference images and local object recognition needs. This customization ensures that the most relevant filters are retained, maintaining high accuracy for locally important objects while achieving faster inference speeds
Solution Approach 2:
The patent performs preliminary analysis using the parent neural network on reference images captured at each edge node to identify which filters are most relevant. This preliminary action guides the pruning process to retain only necessary filters, ensuring accuracy is maintained for locally important objects while enabling faster inference
Data Source
AI summary
A system and method for providing object recognition using artificial neural networks. The method includes capturing a plurality of reference images with a camera associated with an edge node on a communication network. The reference images are received by a centralized server on the communication network. The reference images are analyzed with a parent neural network of the centralized server to determine a subset of objects identified by the parent neural network in the reference images. One or more filters that are responsive to the subset of objects are selected from the parent neural network. A pruned neural network is created from only the one or more filters. The pruned neural network is deployed to the edge node. Real-time images are captured with the camera of the edge node and objects in the real-time images are identified with the pruned neural network.