相关实验视频
Updated: Jan 29, 2026

Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
解释深度神经网络的性能与部分信息分解
Tianyue Liu1,2,3, Binghui Guo1,2,3,4, Ziqiao Yin1,2,3,4,5
1School of Artificial Intelligence, Beihang University, Beijing 100191, China.
深度神经网络 (DNN) 难以应对现实世界的数据转移. 本研究引入了部分信息分解 (PID) 框架,显示DNN中的冗余性更高,协同效应较低,提高了对数据腐败的稳定性.
科学领域:
- 人工智能的人工智能
- 机器学习 机器学习
- 信息理论 信息理论
背景情况:
- 深度神经网络 (DNN) 在面对分布式转移时表现出强度的局限性,阻碍了现实世界的部署.
- 目前对DNN内部信息表示如何与稳定性相关的理解是有限的.
研究的目的:
- 开发一个可解释的框架,用部分信息分解 (PID) 来评估DNN的稳定性.
- 量化冗余性,独特性和协同作用在神经信息编码中对模型稳定性的作用.
主要方法:
- 提出了一个基于部分信息分解 (PID) 的可解释框架.
- 从清洁输入分析了PID措施 (冗余性,唯一性,协同性),以评估神经元编码的信息.
- 与自然腐败下的模型性能相关的PID指标.
主要成果:
- 具有较高冗余率和较低协同率的模型在各种自然腐败下表现更加稳定.
- 更高的独特信息率与对清洁数据的更好的分类准确性有积极的关联.
- 通过内部信息分析证明了轻量级稳定性评估的可行性.
结论:
- 部分信息分解为理解和比较DNN行为提供了新的见解.
- 对内部表示的信息理论分析可以预测并潜在地提高模型稳定性,而无需大量损坏的数据.
更多相关视频
03:31Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
09:01A Method for Investigating Age-related Differences in the Functional Connectivity of Cognitive Control Networks Associated with Dimensional Change Card Sort Performance
Published on: May 7, 2014
相关概念视频
Synthesis and Decomposition Reactions
Protein Networks
These interactions can be represented through maps depicting protein-protein interaction networks, represented as nodes and edges. Nodes are circles that are representative of a protein,...
Partial Fractions
Interpreting R Charts
An R chart plots the range of subsets of measurements collected from a process. Each point on the chart represents the range—defined as the difference between the maximum and minimum...
Network Covalent Solids
To break or to melt a covalent network solid, covalent bonds must be broken. Because covalent bonds are relatively strong, covalent network solids are typically...
Interpreting Run Charts