富里埃或波形基作为对应的自我注意力在spikformer有效的视觉分类的基础
Qingyu Wang1,2, Duzhen Zhang1, Xinyuan Cai1
1Institute of Automation, Chinese Academy of Sciences, Beijing, China.
Frontiers in neuroscience
|February 13, 2025
概括
这项研究介绍了基于福里埃或波纹的尖峰造型器 (FWformer),这是一种节能模型,可以取代尖峰神经网络 (SNN) 中的标准自我注意. 在视觉任务中,FWformer实现了与提高速度和降低能耗的可比准确性.
科学领域:
- 人工智能的人工智能
- 计算神经科学是一种神经科学.
- 信号处理 信号处理
背景情况:
- 与变压器集成的尖端神经网络 (SNN) 提供能源效率和高精度.
- 标准尖峰自我注意力 (SSA) 可能是计算密集的,并且对于稀疏的尖峰数据并不总是必要的.
- 假设SSA和Fourier和Wavelet等变换都使用基础函数来处理信息.
研究的目的:
- 提出和评估一种新的节能尖峰变压器架构.
- 将SSA中的动态基础替换为来自福里埃和波形变换的固定基础.
- 在视觉分类中评估拟议的基于福里埃或波纹的尖峰变频器 (FWformer) 在视觉分类中的性能.
主要方法:
- 开发了基于里埃或波形的尖子变换器 (FWformer),通过用尖子形式的里埃变换,波形变换或它们的组合替换标准的SSA.
- 使用固定三角形或波形基来进行信息转换,而不是动态基.
- 通过使用静态图像和基于事件的视频数据集对视觉分类任务验证了FWformer.
主要成果:
- 与标准的尖子变压器相比,FWformer 实现了同等或更高的精度 (0.4%-1.5%).
- 跑步速度显著改善:9%-51%的训练和19%-70%的推断.
- 报告的理论能耗降低 (20%-25%) 和GPU内存使用率降低 (4%-26%).
结论:
- 该FWformer提供了一个更有效的替代标准spikformer通过利用固定的基础从富里埃和波形变换.
- 这种方法成功地平衡了准确性,大幅提高了速度和降低了计算成本.
- 这些发现表明,通过生物信号 (尖端) 和信息理论 (变形) 启发的改进变压器是一个有前途的研究方向.
相关概念视频
01:10Fast Fourier Transform
258
The Fast Fourier Transform (FFT) is a computational algorithm designed to compute the Discrete Fourier Transform (DFT) efficiently. By breaking down the calculations into smaller, manageable sections, the FFT significantly reduces the computational complexity involved. Direct computation of an N-point DFT requires N2 complex multiplications, whereas the FFT algorithm needs only (N/2)log2N multiplications, offering a much faster performance.
The computational efficiency of the FFT becomes...
The computational efficiency of the FFT becomes...
258
01:15Discrete Fourier Transform
209
The Discrete Fourier Transform (DFT) is a fundamental tool in signal processing, extending the discrete-time Fourier transform by evaluating discrete signals at uniformly spaced frequency intervals. This transformation converts a finite sequence of time-domain samples into frequency components, each representing complex sinusoids ordered by frequency. The DFT translates these sequences into the frequency domain, effectively indicating the magnitude and phase of each frequency component present...
209
01:21Association Areas of the Cortex
4.9K
Association areas are regions of the cerebral cortex that do not have a specific sensory or motor function. Instead, they integrate and interpret information from various sources to enable higher cognitive processes such as memory, learning, and decision-making. Some key association areas include the following:
Prefrontal Association Area: This area is located in the frontal lobe and is involved in planning, decision-making, and moderating social behavior. It connects with primary motor areas,...
Prefrontal Association Area: This area is located in the frontal lobe and is involved in planning, decision-making, and moderating social behavior. It connects with primary motor areas,...
4.9K
01:07Basic signals of Fourier Transform
465
The Fourier Transform is a pivotal mathematical tool in signal processing, enabling the transformation of time-domain signals into their frequency-domain representations. Among the numerous elements within this domain, certain functions like the sinc function, delta function, and exponential signals hold significant importance due to their unique properties and implications.
The sinc function, defined as sinc(x) = sin(πx)/(πx), is particularly notable for its symmetry and behavior at...
The sinc function, defined as sinc(x) = sin(πx)/(πx), is particularly notable for its symmetry and behavior at...
465
01:11Continuous -time Fourier Transform
277
The Fourier series is instrumental in representing periodic functions, offering a powerful method to decompose such functions into a sum of sinusoids. This technique, however, necessitates modification when applied to nonperiodic functions. Consider a pulse-train waveform consisting of a series of rectangular pulses. When these pulses have a finite period, they can be accurately represented by a Fourier series. Yet, as the period approaches infinity, resulting in a single, isolated pulse, the...
277
01:30Classification of Signals
381
In signal processing, signals are classified based on various characteristics: continuous-time versus discrete-time, periodic versus aperiodic, analog versus digital, and causal versus noncausal. Each category highlights distinct properties crucial for understanding and manipulating signals.
A continuous-time signal holds a value at every instant in time, representing information seamlessly. In contrast, a discrete-time signal holds values only at specific moments, often denoted as x(n), where...
A continuous-time signal holds a value at every instant in time, representing information seamlessly. In contrast, a discrete-time signal holds values only at specific moments, often denoted as x(n), where...
381