没有合适的资源？快使用搜索试试~ 我知道了~

首页学习摘要：Methods for interpreting and understanding deep neural networks

学习摘要：Methods for interpreting and understanding deep neural netw...

CNN可视化

需积分: 20 21 下载量 74 浏览量更新于2023-03-16 评论收藏 3.58MB PPTX 举报

身份认证购VIP最低享 7 折!

领优惠券(最高得80元）

试读

50页

关于论文“Methods for interpreting and understanding deep neural networks”的学习摘要

资源详情

资源评论

资源推荐

1. 简介

深度网络主要应用场景：

1 、图像分类

2 、语音识别

3 、自然语言处理

可解释性对与医疗、自动驾驶很重要

因为必须保证模型依赖于正确的特征

本文内容：

ML 方法解释的综述

重点关注 DNN

2. 正文前页

事后可解释性：

例如，给定训练好的模型

目标：理解模型预测的（分类），什么是可解释的（输入变量）

Understanding ：模型的 functional understanding

seek to 刻画模型的黑箱行为

而不是阐明他的内部运作和内部表示

本文关注：

解释 DNN 的输出，并解释单独的预测

定义 1 ：

Interpretation ：是抽象概念（预测的类）到某个人类能理解的 domain 的映射

可解释的 domains 的例子：图像（像素矩阵）、文本（单词序列）

不可解释的例子：抽象的向量空间、无事实证明的输入特征（例 - 无意义的单词序列）

定义 2 ：

Explanation ：是可解释域的特征集合，他们能对于给定一个例子，能帮助产生决策（例 - 分类、回归）

这些组成 explanation 的特征，有相关的 score 来表示每个特征的 contribution

一般 explanation 是真值向量，相关特征是正的，不相关为 0

3. Interpretation DNN 模型

需要解析的通常是 top layer 的表示

因为 top-layer 的神经元是抽象的，且 DNN 输入（图像或文本）通常是可解析的

本部分：

描述如何在输入域内创建 prototype （原型）

这个原型是抽象学习概念的解释器和代表

它可以在 AM 框架中 formulated

3.1 Activation maximization(AM)

AM 是一个分析框架，寻找对于 interest 的，使得模型 input pattern ，使得模型响应最大

可以想象 DNN 分类器，将输入映射到分类集，输出对分类概率编码

类的表示可以用进行最优化

DNN 的类概率模型是带有梯度的函数，所以最优化可以用梯度下降

最右的规范，实现了对接近 origin 的 inputs 的偏好

当应用到图像分类， prototypes 多是灰度图形式，会产生强烈的类响应

3.2 用 expert 提升 AM

为了得到更有意义的 prototype

可以被更复杂的代替，叫做“ expert” ，它可以是模型

最优化问题变成：

通过解决这个最优化问题得到的，将同时产生强烈的类反应，且使得数据类似

应用贝叶斯规则，做误差和常量项目建模

新目标可以被定义为 class-conditioned 数据密度

可能的 expert 是，他可以表示复杂的分布，且在输入域有梯度

：

，在原来 regularizer 的位置

, 从 data 中学习到的

在 interpret 类似自然图像分类，需要更复杂的密度模型，例如卷积 RBM/DBMs 、 pixel-RNNs

Expert 的选择，决定了结果 prototype 的 appearance

剩余49页未读，继续阅读

Write a paper about Deep-learning based analysis of metal-transfer images in GMAW process

The proposed approach can provide insights for process optimization and reduce the need for human expertise in interpreting high-speed images. Future work includes investigating the use of our ...

generic attention-model explainability for interpreting bi-modal and encoder

通用注意力模型是一种用于深度学习模型的解释性方法，它允许我们了解模型对输入的关注点。这种模型可以帮助我们解释为什么神经网络在做出特定预测时会关注于输入的某些部分。通过使用通用注意力模型，我们可以进一步...

This article takes access control and attendance management as the foundation and develops a small IoT project through embedded technology and radio frequency identification technology. For this small project, the RFID module and WIFI module are key components supporting the network communication of the entire project. The RFID module transmits the signal of the card through electromagnetic coupling, while the WIFI module essentially performs wireless transmission through electromagnetic waves, one is active transmission and the other utilizes passive transmission. These two methods, together with the Internet, constitute the large network of the IoT. The IoT transfers data through these two methods and realizes various functions by interpreting the data, making significant contributions to modernization in today's society.翻译为中文

本文以门禁控制和考勤管理为基础，通过嵌入式技术和射频识别技术开发了一个小型物联网项目。在这个小项目中，RFID模块和WIFI模块是支持整个项目网络通信的关键组件。RFID模块通过电磁耦合传输卡片信号，而WIFI模块...

Statistics and Machine Learning Toolbox

The Statistics and Machine Learning ... It provides a powerful and flexible platform for analyzing and interpreting complex data sets, making it an essential tool for researchers and data analysts.

how to do IPTW by R

To do IPTW by R, you can follow these steps: ... It's important to consult with a statistician or other expert to ensure that you are using the appropriate methods and interpreting the results correctly.

material studio

Some of the key features of Material Studio include molecular dynamics simulation, quantum chemistry calculations, crystal structure analysis, and visualization tools for analyzing and interpreting ...

In<!DOCTYPE note SYSTEM’’Note.dtd’’>’,what is the meaning of’note’and what is the meaning of’Note.dtd’’?

In `<!DOCTYPE note SYSTEM "Note.dtd">`, "note" refers to the name of the ... It serves as a blueprint for validating and interpreting the content of documents that conform to the "note" document type.

gesture识别长划短划

Gesture recognition is the process of understanding and interpreting human gestures or movements. A long swipe or stroke is typically associated with a continuous or extended movement, while a short ...

[Errno 22] Invalid argument: 'E:\\pyj\x0ben\\pyz\x03_film.csv'

3. Use raw string notation to avoid interpreting escape characters in the file path. For example, instead of 'E:\\pyj\x0ben\\pyz\x03_film.csv', use r'E:\pyj\x0ben\pyz\x03_film.csv'. 4. Ensure that the...

SyntaxError: (unicode error) 'unicodeescape' codec can't decode bytes in position 2-3: truncated \UXXXXXXXX escape

This error occurs when trying to ...Both of these solutions will prevent Python from interpreting the backslash as an escape character and will allow the Unicode escape sequence to be properly decoded.

SyntaxError: (unicode error) 'unicodeescape' codec can't decode bytes in position 4-5: truncated \UXXXXXXXX escape

To fix this error, you should check the string for any incomplete or invalid escape sequences and correct them. You can also try using a raw string by adding an 'r' before the string to avoid ...

scala> :paste // Entering paste mode (ctrl-D to finish) from pyspark.sql import SparkSession // Exiting paste mode, now interpreting. <pastie>:1: error: ';' expected but '.' found. from pyspark.sql import SparkSession

这个错误通常是因为Scala代码中使用了Python语法。请注意，Scala和Python语法是不同的，不能混用。如果您想在Scala中使用SparkSession，请按照以下方式导入： ```scala import org.apache.spark.sql.SparkSession ...

'pom.xml' has syntax errors

If you are seeing syntax errors in this file, it means that there is an issue with the XML structure or syntax that is preventing Maven from properly reading and interpreting the file. To fix this ...

闭环反馈卡尔曼滤波器matlab用法

“Estimating and Interpreting the instantaneous frequency of a signal”. In: Proceedings of the IEEE 80.4 (1992), pp. 540–568. - S. Savaresi et al. “Frequency estimation of narrow band signals in ...

cell type aware

This may involve analyzing and interpreting data in a way that distinguishes between different cell types, using techniques that selectively target certain cell populations, or designing experiments ...

<optional>true</optional>

This tag indicates that a particular element or attribute is optional, meaning it can be omitted or ... This information is helpful to developers and users who are creating or interpreting the document.

iyangdi

粉丝: 320
资源: 2

上传资源快速赚钱

我的内容管理收起

我的资源快来上传第一个资源

我的收益

登录查看自己的收益

我的积分登录查看自己的积分

我的C币登录后查看C币余额

我的收藏

我的下载

下载帮助

会员权益专享

学习摘要：Methods for interpreting and understanding deep neural netw...

评论0

会员权益专享

最新资源

学习摘要：Methods for interpreting and understanding deep neural netw...

评论0

mnist-explanation:在MNIST数据集上试用DNN的复制技术

A Peek Into the Reasoning of Neural Networks Interpreting With

An Introduction to Deep Learning for the Physical Layer

给出10个深度学习课程设计的题目

Write a paper about Deep-learning based analysis of metal-transfer images in GMAW process

generic attention-model explainability for interpreting bi-modal and encoder

Statistics and Machine Learning Toolbox

how to do IPTW by R

material studio

In<!DOCTYPE note SYSTEM’’Note.dtd’’>’,what is the meaning of’note’and what is the meaning of’Note.dtd’’?

gesture识别长划短划

[Errno 22] Invalid argument: 'E:\\pyj\x0ben\\pyz\x03_film.csv'

SyntaxError: (unicode error) 'unicodeescape' codec can't decode bytes in position 2-3: truncated \UXXXXXXXX escape

SyntaxError: (unicode error) 'unicodeescape' codec can't decode bytes in position 4-5: truncated \UXXXXXXXX escape

scala> :paste // Entering paste mode (ctrl-D to finish) from pyspark.sql import SparkSession // Exiting paste mode, now interpreting. <pastie>:1: error: ';' expected but '.' found. from pyspark.sql import SparkSession

'pom.xml' has syntax errors

闭环反馈卡尔曼滤波器matlab用法

cell type aware

<optional>true</optional>

会员权益专享

最新资源