<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[baibaidoudou的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/baibaidoudou</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; baibaidoudou]]></copyright><item><title><![CDATA[Deep Reinforcement Learning]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119758245</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119758245</guid><author>baibaidoudou</author><pubDate>Fri, 27 Aug 2021 16:30:52 +0800</pubDate><description><![CDATA[当你发现搜集有label的资料很困难的时候，正确答案也不知道的时候，可以考虑用RL
]]></description><category></category></item><item><title><![CDATA[Explainable Machine Learning]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119534204</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119534204</guid><author>baibaidoudou</author><pubDate>Mon, 09 Aug 2021 17:27:58 +0800</pubDate><description><![CDATA[1.定义   

Explainable指的是有一个东西是黑箱，我们想办法赋予它解释的能力

Interpretable指的是有一个东西本来就不是黑箱



2.Goal of Explainable ML

Make people comfortable ，给人们一个理由让人们更好接收





3.Local Explanation

Which component is critical



Removing or modifying the componets

large decisi...]]></description><category></category></item><item><title><![CDATA[Adversarial Attack]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119510010</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119510010</guid><author>baibaidoudou</author><pubDate>Sun, 08 Aug 2021 22:17:18 +0800</pubDate><description><![CDATA[Adversarial Attack]]></description><category></category></item><item><title><![CDATA[Self-supervised Learning——BERT]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119418732</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119418732</guid><author>baibaidoudou</author><pubDate>Thu, 05 Aug 2021 16:48:26 +0800</pubDate><description><![CDATA[1.Self-supervised Learning



Supervised是要有label的资料

Self-supervised是让资料一部分作为model，一部分作为label



2.Masking Input



两种方法，要么随机盖住某个token，要么将此token随机换成别的词。



对其做Linear transform（矩阵）然后进行softmax得到一个分布



我们知道盖住的词是什么，BERT不知道，所以看输出的词与原来词的最小交叉熵



3.How to us...]]></description><category></category></item><item><title><![CDATA[GAN(3)]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119414525</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119414525</guid><author>baibaidoudou</author><pubDate>Thu, 05 Aug 2021 12:35:16 +0800</pubDate><description><![CDATA[1.Conditional GAN

给个条件x





Text-to-Image



要让discriminator给高分，x与y即文字与影像要是相配的





2.Learning from Unpaired Data



使用Cycle GAN












]]></description><category></category></item><item><title><![CDATA[GAN（2）]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119350122</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119350122</guid><author>baibaidoudou</author><pubDate>Thu, 05 Aug 2021 11:46:20 +0800</pubDate><description><![CDATA[1.Our Ojective

找到一个类似于loss function的东西



Normal Dsitribution经过G(generator)后得到一个分布，data有一个分布，我们的目的是找一组Generator的参数让两个P更相似，写作。



2.Solution

GAN（生成对抗网络）就是找到上文的方法。

不需要知道,的分布，只需要从它们中sample（抽样）一些。

这里就用到了Discriminator



的给高分，给低分，使最大



3.Wasserstein di...]]></description><category></category></item><item><title><![CDATA[GAN（生成式对抗网络）]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119278314</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119278314</guid><author>baibaidoudou</author><pubDate>Sat, 31 Jul 2021 17:31:29 +0800</pubDate><description><![CDATA[1. Unconditional generation（无条件生成）



generator的作用是产生一个高维向量



2.Discriminator（辨别器）



输出是个数值（scalar），数字越大越是想得到的东西



3.Algorithm   

Initialize generator and discriminator：G &amp; D

Step1：Fix generator G, and update discriminator D

Discriminato...]]></description><category></category></item><item><title><![CDATA[Transformer]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119246048</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119246048</guid><author>baibaidoudou</author><pubDate>Fri, 30 Jul 2021 17:04:55 +0800</pubDate><description><![CDATA[1.Seq2seq






]]></description><category></category></item><item><title><![CDATA[Self-attention]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/119213733</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/119213733</guid><author>baibaidoudou</author><pubDate>Thu, 29 Jul 2021 16:30:31 +0800</pubDate><description><![CDATA[1.运行机制



每个output考虑所有input的vector

Find the relevant vectors in a sequence





2.计算模组

1.Dot - product（最常用方法）



其中W矩阵通过计算得到

2.Additive







3.计算关联度（）



softmax =

并不一定用softmax







4.Muti-head Self-attention



5.Positional Encoding



自...]]></description><category></category></item><item><title><![CDATA[机器学习笔记-CNN]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118703414</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118703414</guid><author>baibaidoudou</author><pubDate>Tue, 13 Jul 2021 17:31:33 +0800</pubDate><description><![CDATA[1.Image Classification

一张图片可以分成三个维度（RGB)



展开成100 * 100 * 3的向量，作为输入

 Fully Connected Network

1.1 Observation 1



根据不同的特征，来判断一个物件

因此不需要把整个图片作为输入！

1.1.2Simplification 1

filter滤波器的思想？

Receptive field : 感受野(自己决定）

1.1.3Simplification 1 - Typical...]]></description><category></category></item><item><title><![CDATA[Normalization]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118698649</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118698649</guid><author>baibaidoudou</author><pubDate>Tue, 13 Jul 2021 15:03:11 +0800</pubDate><description><![CDATA[1. 什么是Feature Normalization





standard deviation：标准差

variances： 方差

取出同一纬度i的所有x，计算它们的平均值记为

根据公式求得放入这一维度中，这样他们的平均值为0，方差都是1，方便做gradient descent.



2. Considering Deep Learning





3.Batch Normalization- Testing



We do not always have batch at t...]]></description><category></category></item><item><title><![CDATA[Classification]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118697790</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118697790</guid><author>baibaidoudou</author><pubDate>Tue, 13 Jul 2021 13:34:29 +0800</pubDate><description><![CDATA[1.Classification as Regression?

Regression:



Classification as regression

 class

1 = class1， 2 = class2， 3 = class3

希望输出的y与class编号越接近越好



2.Soft-max

             

 









3.Loss of Classification



Mean Square Error(MSE) :...]]></description><category></category></item><item><title><![CDATA[Optimization]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118675575</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118675575</guid><author>baibaidoudou</author><pubDate>Mon, 12 Jul 2021 16:52:12 +0800</pubDate><description><![CDATA[Reasons forfailure

1.gradient is close to zero，卡在了critical point(驻点）

判断critical point是不是saddle point，求出H



Sometimes,sometimes  Saddle point

is the eigen value of H,is an eigen vector of H

如果卡在了Saddle point，找出为负的时候，所对应的，只要顺着的方向去更新参数，就可以找到更低...]]></description><category></category></item><item><title><![CDATA[机器学习任务攻略]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118655527</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118655527</guid><author>baibaidoudou</author><pubDate>Sun, 11 Jul 2021 17:33:48 +0800</pubDate><description><![CDATA[1. 检查loss on training data

2.如果loss大，检查bias(是不是model太过简单)

            optimization issue：没有合适的方法（除了梯度下降法）

 如果loss小，检查loss on testing data

3.如果大，考虑overfitting（解决方法：1.增加训练集，2.添加模型限制）

         mismatch（training和testing data 有不一样的分...]]></description><category></category></item><item><title><![CDATA[Deep learning]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118651440</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118651440</guid><author>baibaidoudou</author><pubDate>Sun, 11 Jul 2021 11:05:32 +0800</pubDate><description><![CDATA[Fully Connect Feedforward Network 全连接前向传播网络

cross entropy：交叉熵



Backpropagation：反向传播

Chain Rule：链式法则

1.Forward pass

Computefor all parameters





 2. Backward pass

Computefor all activation function(激活函数)inputs z



假设,知道了





现在问题是如何计算,
...]]></description><category></category></item><item><title><![CDATA[第一节-机器学习基本概念]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118637104</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118637104</guid><author>baibaidoudou</author><pubDate>Sat, 10 Jul 2021 20:27:53 +0800</pubDate><description><![CDATA[machine learning ≈ looking for function

Different types of functions:

Regression(回归): The function outputs a scalar

Classification(分类): Given options(classes),the function outputs the correct one

structured learning(产生结构，创造）



How to find a function?
]]></description><category></category></item><item><title><![CDATA[（python学习日记）2.列表]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/118345802</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/118345802</guid><author>baibaidoudou</author><pubDate>Wed, 30 Jun 2021 14:20:14 +0800</pubDate><description><![CDATA[2.1 列表是什么


bicycle=['trek','cannondale','readline','specialized']
print(bicycle)

2.2 访问列表元素


bicycle=['trek','cannondale','readline','specialized']
print(bicycle[0])
bicycle=['trek','cannondale','readline','specialized']
print(bicycle[1])
bicycle=['trek]]></description><category></category></item><item><title><![CDATA[（python学习日记）1.变量]]></title><link>https://blog.youkuaiyun.com/baibaidoudou/article/details/117658244</link><guid>https://blog.youkuaiyun.com/baibaidoudou/article/details/117658244</guid><author>baibaidoudou</author><pubDate>Tue, 29 Jun 2021 22:53:49 +0800</pubDate><description><![CDATA[message="Hellow World"
print(message)
message="David"
print(message)

变量的命名与使用


#字母，数字，下划线
message="xxxxxx"
1_message="xxxx"(x)
_messge="xxxx"
0="xxxxx"(x)
#打x的都是不能这样命名的


]]></description><category></category></item></channel></rss>