相关实验视频
Updated: Mar 6, 2026
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
10.1K
推进前进算法,实现高性能深度本地学习
Siyuan Xu1, Yujie Wu2, Jibin Wu3
1Department of Computing, The Hong Kong Polytechnic University, China; Institute of Automation, Key Laboratory of Brain Cognition and Brain-inspired Intelligence Technology, Chinese Academy of Sciences, Beijing, China; University of Chinese Academy of Sciences, Beijing, China.
概括
一个新的距离前进 (DF) 算法增强了神经网络的前进前进 (FF) 方法. 这种方法提高了性能和概括性,同时保持了内存效率和并行化优势,而不是反向传播.
科学领域:
- 人工智能的人工智能
- 机器学习 机器学习
- 计算神经科学是一种神经科学.
背景情况:
- 反向传播 (BP) 在内存效率和并行化方面面临限制.
- 前进前进 (FF) 算法提供了本地学习替代方案,但在性能和通用化方面存在困难.
- 现有的FF方法缺乏有效的学习策略来处理复杂的任务.
研究的目的:
- 为了提高FF算法的性能和一般化在监督学习中.
- 为了保持FF的有利的本地计算特性.
- 将基于FF的本地学习扩展到尖端神经网络 (SNN) 和神经形态硬件.
主要方法:
- 通过使用远程度量学习重构的FF,提出了远程前 (DF) 算法.
- 开发了一种基于善良的N-pair边际损失,用于歧视性特征学习.
- 集成层级协作本地更新策略以减轻信息丢失.
- 扩展了DF到SNN,用于时间尖峰序列的好性函数.
主要成果:
- 在八个数据集中,DF算法超越了现有的FF模型和本地学习方法.
- 与BP训练相比,DF方法实现了超过60%的内存成本降低.
- 拟议的方法证明了对与硬件相关的噪声的强化稳定性.
- 在神经形态硬件上进行事件驱动处理的有效实现.
结论:
- 拟议的DF算法提供了一个高效和强大的本地学习解决方案.
- 在保持内存效率的同时,DF提高了FF的性能和通用性.
- 这种方法对未来的FF算法设计和多核硬件应用具有前景.
相关概念视频
01:29Implicit Differentiation: Problem Solving
83
Curves defined implicitly, where variables cannot be separated algebraically, require specialized techniques for analysis. The conchoid of Nicomedes exemplifies such a case. Its equation links x and y in a way that prevents isolation of one variable, making implicit differentiation essential to determine the slope and behavior at any point on the curve.The implicit form of the conchoid can be expressed as:To differentiate this equation, y is treated as a function of x, and the chain rule is...
83
01:25First Derivative Test: Problem Solving
102
Imagine an asset price that crashes to a low point, rebounds sharply as bargain-hunters step in, and then gradually declines. Such behavior can be modeled with a smooth function whose turning points represent locally overvalued and undervalued regions. A convenient example that captures rebound followed by decay is:The high and low points of this curve are identified using the first derivative test, which determines where the function changes from increasing to decreasing or vice versa. To...
102