Evolutionary Dual-Stream Transformer.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Vision Transformers (ViTs) are powerful but computationally intensive, hindering their widespread application.
• High computational requirements of current ViTs limit advancements in computer vision research and development.

Purpose of the Study:

• To propose a novel Evolutionary Dual-Stream Transformer (E-DST) model to address the computational demands of Vision Transformers.
• To develop an efficient model that fuses convolutional and transformer features for computer vision tasks.

Main Methods:

• Introduced a hybrid attention mechanism within a Dual-Stream Transformer (DST) architecture.
• Employed a dual-branch structure to integrate convolutional and transformer-based feature learning.
• Developed and proved the convergence of a novel evolutionary optimizer for model parameter optimization.

Main Results:

• The E-DST model demonstrated significant reduction in computational resource requirements compared to classic models.
• The evolutionary optimizer proved effective in optimizing transformer parameters and showed generality across various neural network architectures.
• Experimental validation on three datasets confirmed the E-DST model's effectiveness and efficiency.

Conclusions:

• The proposed E-DST model offers a feasible and effective solution for reducing computational costs in computer vision.
• The evolutionary optimizer is capable of solving large-scale optimization problems inherent in deep learning models.
• This research advances the efficiency and applicability of transformer-based models in computer vision.