重新思考视觉转换器中的注意力机制与图形结构.
Hyeongjin Kim1, Byoung Chul Ko1
1Department of Computer Engineering, Keimyung University, Daegu 42601, Republic of Korea.
Sensors (Basel, Switzerland)
|February 24, 2024
概括
本研究介绍了图形头部注意力视觉转换器 (GHA-ViT),通过维护本地和全球补丁信息来改进图像分析. 与标准视觉变压器相比,GHA-ViT提高了性能并降低了参数.
科学领域:
- 计算机视觉 计算机视觉
- 机器学习 机器学习
- 人工智能的人工智能
背景情况:
- 标准视觉变压器 (ViT) 使用多头注意力 (MHA),这是参数密集型的,可以损害图像局部性.
- 需要更高效和有效的ViT架构来保护空间信息.
研究的目的:
- 提出一种新的视觉变压器架构,GHA-ViT,结合图形头部注意力 (GHA).
- 提高ViT的性能,同时降低计算复杂性和参数数量.
主要方法:
- 在标准ViT中取代了多头注意力 (MHA) 机制,采用了新的图形头注意力 (GHA).
- 将图形结构应用于变压器的注意力头,以更好地捕捉图像补丁中的关系.
- 在各种数据集上评估了GHA-ViT,包括CIFAR-10/100,MNIST,MNIST-F和ImageNet-1K.
主要成果:
- 在多个数据集中,GHA-ViT表现出了比纯ViT模型更高的性能.
- 在ImageNet-1K上使用GHA-B模型 (约1.7%) 实现了81.7%的Top-1精度. 29M参数). 这就是为什么.
- 与现有的ViT相比,在CIFAR-10/100上显著减少参数 (17倍) 和提高性能 (0.4%/4.3%).
结论:
- 拟议的GHA-ViT有效地保持了图像补丁的本地性和全球性,确保了注意力的多样性.
- GHA-ViT为当前最先进的ViT模型提供了一个有希望的轻量级替代方案,平衡精度,参数数量和计算操作.
相关概念视频
Visual System
582
Light enters the eye through the cornea, a transparent, dome-shaped surface covering the surface of the eyeball that helps to direct and focus incoming light. This light is then channeled toward the pupil, an adjustable opening whose size is controlled by the iris. The iris, a pigmented muscle, regulates the amount of light entering the eye by contracting or dilating the pupil, thereby ensuring optimal light levels for clear vision.
Once through the pupil, the light passes through the lens, a...
Once through the pupil, the light passes through the lens, a...
582
Vision
53.3K
Vision is the result of light being detected and transduced into neural signals by the retina of the eye. This information is then further analyzed and interpreted by the brain. First, light enters the front of the eye and is focused by the cornea and lens onto the retina—a thin sheet of neural tissue lining the back of the eye. Because of refraction through the convex lens of the eye, images are projected onto the retina upside-down and reversed.
53.3K
Color Vision
578
Color perception begins in the retina, the light-sensitive layer at the back of the eye. Two main theories explain how colors are seen: the trichromatic theory and the opponent-process theory. The trichromatic theory, proposed by Thomas Young in 1802 and extended by Hermann von Helmholtz in 1852, suggests that color vision is based on three types of cone receptors in the retina. These cones are sensitive to different but overlapping ranges of wavelengths corresponding to red, blue, and green.
578
Anatomy of the Eyeball
7.1K
The eye is a spherical, hollow structure composed of three tissue layers. The outer layer — the fibrous tunic, comprises the sclera — a white structure — and the cornea, which is transparent. The sclera encompasses some of the ocular surface, most of which is not visible. However, the 'white of the eye' is distinctively visible in humans compared to other species. The cornea, a clear covering at the front of the eye, enables light penetration. The eye's middle...
7.1K
Parallel Processing
151
The brain processes sensory information rapidly due to parallel processing, which involves sending data across multiple neural pathways at the same time. This method allows the brain to manage various sensory qualities, such as shapes, colors, movements, and locations, all concurrently. For instance, when observing a forest landscape, the brain simultaneously processes the movement of leaves, the shapes of trees, the depth between them, and the various shades of green. This enables a quick and...
151
The Retina
69.1K
The retina is a layer of nervous tissue at the back of the eye that transduces light into neural signals. This process, called phototransduction, is carried out by rod and cone photoreceptor cells in the back of the retina.
69.1K


