Related Experiment Video
Updated: Aug 12, 2026
13:51
Cross-Modal Multivariate Pattern Analysis
Published on: November 9, 2011
BRAINHash: Brain-inspired Region-Aligned Interaction Network for unsupervised cross-modal hashing
Summary
BRAINHash introduces a novel brain-inspired approach for unsupervised cross-modal hashing, overcoming limitations in sample pairing and network capacity. This method enhances retrieval accuracy by mimicking cognitive functions for more robust cross-modal data representation.
Area of Science:
- Artificial Intelligence
- Computer Science
- Computational Neuroscience
Background:
- Unsupervised cross-modal hashing (UCMH) methods face challenges in unsupervised sample pairing and static network architectures.
- Existing contrastive learning frameworks with MLPs struggle with reliable positive-negative sample construction.
Purpose of the Study:
- To propose BRAINHash, a novel brain-inspired memory-based temporal modeling strategy for UCMH.
- To address limitations in unsupervised sample pairing and representational capacity of existing UCMH methods.
Main Methods:
- BRAINHash employs modular design mimicking brain functions: diverse encoders (occipital lobe), error-aware optimization (prefrontal cortex), teacher network for memory (hippocampus), and SNN-based temporal modeling.
- Integrates biologically inspired architectural principles with temporal dynamics for UCMH.
Main Results:
- BRAINHash outperforms fifteen state-of-the-art approaches on five widely-used datasets.
- Demonstrates superior performance in unsupervised cross-modal hashing tasks.
Conclusions:
- BRAINHash represents a pioneering integration of brain-inspired design and temporal dynamics in UCMH.
- The proposed method offers a significant advancement in unsupervised cross-modal retrieval systems.
