Artificial Neural Network with Long-Range Horizontal Connections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial neural networks require a deep structure with a large number of layers for object recognition, leading to high calculation and energy consumption, resulting in low connection efficiency compared to performance, similar to the human visual cortex structure.
Innovation Solution
An electronic device utilizing an artificial neural network with long-range horizontal connections, comprising a small number of layers and neurons, to achieve efficient object recognition by optimizing the ratio of long-range to local connections, allowing the network to operate similarly to the animal's visual cortex structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep structure with a large number of layers is used in the artificial neural network, then object recognition performance is improved, but calculation amount and energy consumption increase significantly
Solution Approach 1:
The patent introduces long-range horizontal connections that span across multiple layers and spatial distances, adding a dimensional aspect to information flow that bypasses traditional sequential layer-by-layer processing. This allows the network to capture global contextual information without increasing depth, thereby maintaining recognition performance while reducing the need for excessively deep structures and associated energy consumption.
Solution Approach 2:
The long-range horizontal connection module serves multiple functions simultaneously: it provides global contextual information, enables skip connections for feature preservation, and facilitates information sharing across different spatial locations. This multi-functionality allows the network to achieve high recognition performance without requiring separate dedicated structures for each function, thus reducing overall computational burden and energy consumption.
2Measurement precision
If a deep structure with a large number of layers is used in the artificial neural network, then object recognition performance is improved, but connection efficiency decreases
Solution Approach 1:
By introducing horizontal connections that operate across spatial dimensions and layer depths simultaneously, the patent creates a multi-dimensional information flow architecture. This allows the network to maintain high connection efficiency by providing direct shortcuts for information propagation, avoiding the need for excessively deep sequential processing paths that would otherwise be required to achieve the same recognition performance.
3Use of energy by moving object
If a small number of layers is used in the artificial neural network, then energy consumption and calculation amount are reduced, but object recognition performance decreases
Solution Approach 1:
The long-range horizontal connections introduce an additional dimensional pathway for information flow that compensates for the reduced depth of the network. By enabling direct long-distance information propagation across spatial and layer dimensions, the network can maintain high recognition performance with fewer layers, thereby reducing energy consumption and computational requirements.
Solution Approach 2:
The long-range horizontal connections act as intermediary pathways that facilitate direct information exchange between distant neurons and layers. This intermediary mechanism allows shallow network structures to access global contextual information that would otherwise require many intermediate processing layers, thus maintaining recognition performance while reducing overall network depth and energy consumption.
Data Source
AI summary
Various example embodiments relate to an electronic device for resource-efficient object recognition using an artificial neural network with long-range horizontal connections and an operating method thereof, and the artificial neural network is configured to recognize an object from an image, be composed of a plurality of neurons, and include at least one hidden layer including at least one long-range horizontal connection connecting any two of the neurons with a length exceeding a preset distance, and at least one local connection connecting any two of the neurons with a length below a preset distance.