通过影响功能解释CNN的决定
Aisha Aamir1, Minija Tamosiunaite1,2, Florentin Wörgötter1
1Third Institute of Physics - Biophysics and Bernstein Center for Computational Neuroscience, University of Göttingen, Göttingen, Germany.
这项研究揭示了图像干扰如何影响深度学习模型. 层级影响分析有助于识别有影响力的训练数据,提高模型的解释性,减少神经网络中的分类偏差.
科学领域:
- 人工智能的人工智能
- 计算机视觉 计算机视觉
- 机器学习 机器学习
背景情况:
- 深度神经网络的解释性对于理解模型决策和减轻预测偏差至关重要.
- 调试模型输出需要分析它们与训练数据集的关系.
研究的目的:
- 通过分析损失函数扰动来理解深度学习模型行为,特别是类预测.
- 调查图像干扰如何影响模型预测,并确定有影响力的训练数据.
主要方法:
- 计算了VGG16网络的不同隐藏层的影响力得分.
- 引入了三种类型的图像干扰:在ImageNet数据集图像上消除纹理,风格和背景.
- 利用全球和层级影响评分来识别影响测试集预测的关键训练图像.
主要成果:
- 影响力得分成功地确定了给定测试集的有影响力的训练图像.
- 突出了特定的干扰 (纹理,风格,背景) 如何影响网络预测.
- 层对层的影响分析,结合沙普利值等方法,揭示了扰乱图像子组之间的显著差异.
结论:
- 层层的可解释性在预训练的卷积神经网络中有效地识别分类偏差.
- 这种方法为重新训练特定隐藏层提供了有价值的见解,以提高模型的稳定性.
- 这些发现支持使用影响力评分来调试和改进深度学习模型的可解释性.
更多相关视频
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
10:44Inherent Dynamics Visualizer, an Interactive Application for Evaluating and Visualizing Outputs from a Gene Regulatory Network Inference Pipeline
Published on: December 7, 2021
相关概念视频
International Nursing Organizations I
ICN member organizations work to advance the field of nursing and healthcare via policies, partnerships, lobbying, professional...
The Role of Ion Channels in Neuronal Computation
Sometimes a single EPSP is strong enough to induce an action potential in the postsynaptic neuron. However, multiple presynaptic inputs must often create EPSPs around the same time for the postsynaptic neuron to be sufficiently depolarized to fire an action potential....
Conformity
Decision Making: Traditional Method
First, a specific claim about the population parameter is decided based on the research question and is stated in a simple form. Further, an opposing statement to this claim is also stated. These statements can act as null and alternative hypotheses, out of which a null hypothesis would be a...
Decision Making
Automatic decision-making is fast, intuitive, and relies on gut feelings...
Cause and Effect
