Hierarchical Neural Network with Parallel Layers for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network architectures face challenges in reducing the number of weighting coefficients, which hinders processing speed and memory efficiency, especially when trying to maintain recognition accuracy by separating interconnected and non-interconnected layers at different stages.
Innovation Solution
A hierarchical neural network architecture with parallel first and second neural networks, where specific portions of channels in each layer are referenced to calculate feature data, reducing the number of connections and weighting coefficients while maintaining recognition accuracy by allowing connections between neuron groups.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the number of weighting coefficients is reduced to improve processing speed and memory efficiency, then processing speed and memory efficiency are improved, but recognition accuracy deteriorates
Solution Approach 1:
The neural network is divided into multiple layers with different connection patterns. Specifically, the network includes an interconnected layer where neurons are connected to all neurons in the next layer, and non-interconnected layers where neurons are connected only to specific neurons in the next layer. This segmentation allows different parts of the network to have different density characteristics, reducing the total number of weighting coefficients while preserving important connections for accuracy.
Solution Approach 2:
Different connection patterns are applied to different layers of the neural network. The interconnected layer maintains full connectivity to preserve recognition accuracy, while non-interconnected layers use sparse connectivity to reduce the number of weighting coefficients. This local differentiation optimizes the balance between processing speed/memory efficiency and recognition accuracy at different stages of the network.
2Quantity of substance
If the number of weighting coefficients is reduced to reduce memory requirements, then memory requirements are reduced, but processing speed deteriorates
Solution Approach 1:
The neural network architecture segments connections into interconnected and non-interconnected layers, reducing the total number of weighting coefficients that must be stored in memory. This segmentation directly addresses the memory requirement issue while maintaining sufficient connectivity for processing speed through the interconnected layers.
Solution Approach 2:
The connection pattern parameter is changed from uniform full connectivity to a mixed pattern with varying connectivity densities across layers. This parameter change reduces the total number of weighting coefficients (memory requirements) while preserving critical paths for processing speed through the interconnected layer design.
3Measurement precision
If interconnected and non-interconnected layers are separated at different stages to maintain recognition accuracy, then recognition accuracy is maintained, but device complexity increases
Solution Approach 1:
The network is segmented into distinct interconnected and non-interconnected layers with clearly defined connection patterns. This segmentation, while increasing architectural complexity, enables the maintenance of recognition accuracy by preserving full connectivity where needed while reducing overall complexity compared to a fully interconnected network.
Solution Approach 2:
The neural network architecture serves multiple functions through its layered structure: the interconnected layer maintains recognition accuracy through full connectivity, while non-interconnected layers reduce complexity and computational load. This multi-functional design allows the single network to achieve both accuracy and efficiency goals.
Data Source
AI summary
There is provided with a data processing apparatus for detecting an object from an image using a hierarchical neural network. The data processing apparatus has parallel first and second neural networks. An obtaining unit obtains a table which defines different first and second portions. An operation unit performs calculation of the feature data of a third portion based on feature data of the first portion identified using the table and on a weighting parameter between first and second layers of the first neural network, and calculation of feature data of a fourth portion based on feature data of the second portion identified using the table and on a weighting parameter between the first and second layers of the second neural network.