Related Experiment Video
Updated: Nov 26, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
823
A brain-inspired network architecture for cost-efficient object recognition in shallow hierarchical neural networks
Youngjin Park1, Seungdae Baek1, Se-Bum Paik2
1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea.
Summary
Long-range horizontal connections (LRCs) in shallow neural networks improve visual object recognition efficiency, mimicking brain capabilities. These connections enhance performance comparable to deeper networks, suggesting a brain-inspired design for parsimonious recognition.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
- Computer Vision
Background:
- The brain achieves efficient visual object recognition using shallow hierarchical networks.
- Artificial deep neural networks (DNNs) are typically much deeper than biological visual networks.
- The role of long-range horizontal connections (LRCs) in biological vision and their potential in artificial networks is under-explored.
Purpose of the Study:
- To investigate how long-range horizontal connections (LRCs) enable cost-efficient visual object recognition in shallow neural networks.
- To compare the performance of shallow networks with and without LRCs to deeper networks.
- To explore the emergence and necessity of LRCs in network architectures.
Main Methods:
- Simulated a model hierarchical neural network with convergent feedforward connections and LRCs.
- Employed network pruning with gradient-based optimization to study the emergence of LRCs.
- Performed ablation studies by removing emerged LRCs to assess their impact on performance.
Main Results:
- Adding LRCs to shallow feedforward networks significantly enhanced image classification performance, reaching levels comparable to much deeper networks.
- A combination of sparse LRCs and dense local connections maximized performance per wiring cost.
- LRCs emerged spontaneously through optimization for minimal connection length while maintaining performance, and their ablation significantly reduced classification accuracy.
Conclusions:
- LRCs are a crucial component for enabling cost-efficient visual object recognition in shallow neural networks.
- Brain-inspired strategies incorporating LRCs offer a promising approach for designing parsimonious network architectures for object recognition under physical constraints.
- This study provides evidence for the functional importance of LRCs in biological visual systems and their potential application in artificial intelligence.
