当前位置:网站首页>5.过拟合,dropout,正则化
5.过拟合,dropout,正则化
2022-07-07 23:11:00 【booze-J】
过拟合


过拟合导致测试误差变大:
可以看到图中随着模型结构的越来越复杂,训练集的误差越来越小,测试集的误差先变小后变大,过拟合导致测试误差变大。
比较好的情况是训练误差和测试误差这两条线比较接近。
防止过拟合
1.增大数据集
数据挖掘领域流行着这样一句话,“有时候拥有更多的数据胜过一个好的模型”。一般来说更多的数据参与训练,训练得到的模型就越好。如果数据太少,而我们构建的神经网络又太复杂的话就比较容易产生过拟合的现象。
2.Early stopping
在训练模型的时候,我们往往会设置一个比较大的选代次数。Early stopping便是一种提前结束训练的策略用来防止过拟合。
一般的做法是记录到目前为止最好的validation accuracy,当连续10个Epoch没有达到最佳accuracy时,则可以认为accuracy不再提高了。此时便可以停止迭代了(Early Stopping)。
3.Dropout

每次训练的时候,都会随机的去关闭一些神经元,关闭的意思并不是去掉,而是这些画虚线的神经元不参与训练。注意一般训练完,测试模型的时候,是使用所有神经元,不会进行dropout。
4.正则化
C0代表原始的代价函数,n代表样本的个数, λ \lambda λ就是正则项系数,权衡正则项与C0项的比重。
L1正则化:

L1正则化可以达到模型参数稀疏化的效果。
L2正则化:
L2正则化可以使得模型的权值衰减,使模型参数值都接近于0。

当 λ \lambda λ=0.001时,出现了过拟合现象,当 λ \lambda λ=0.01时,有较轻微的过拟合,当 λ \lambda λ=0.1的时候没有出现过拟合现象。
边栏推荐
- Langchao Yunxi distributed database tracing (II) -- source code analysis
- 他们齐聚 2022 ECUG Con,只为「中国技术力量」
- Deep dive kotlin synergy (XXII): flow treatment
- 玩轉Sonar
- DNS 系列(一):为什么更新了 DNS 记录不生效?
- Interface test advanced interface script use - apipost (pre / post execution script)
- The weight of the product page of the second level classification is low. What if it is not included?
- What is load balancing? How does DNS achieve load balancing?
- Installation and configuration of sublime Text3
- 赞!idea 如何单窗口打开多个项目?
猜你喜欢

They gathered at the 2022 ecug con just for "China's technological power"

国外众测之密码找回漏洞

基于卷积神经网络的恶意软件检测方法

大数据开源项目,一站式全自动化全生命周期运维管家ChengYing(承影)走向何方?

Installation and configuration of sublime Text3

Where is the big data open source project, one-stop fully automated full life cycle operation and maintenance steward Chengying (background)?

Codeforces Round #804 (Div. 2)(A~D)

QT establish signal slots between different classes and transfer parameters

FOFA-攻防挑战记录

赞!idea 如何单窗口打开多个项目?
随机推荐
[Yugong series] go teaching course 006 in July 2022 - automatic derivation of types and input and output
Lecture 1: the entry node of the link in the linked list
Where is the big data open source project, one-stop fully automated full life cycle operation and maintenance steward Chengying (background)?
Hotel
A network composed of three convolution layers completes the image classification task of cifar10 data set
Four stages of sand table deduction in attack and defense drill
【obs】Impossible to find entrance point CreateDirect3D11DeviceFromDXGIDevice
Cause analysis and solution of too laggy page of [test interview questions]
新库上线 | 中国记者信息数据
深潜Kotlin协程(二十二):Flow的处理
【愚公系列】2022年7月 Go教学课程 006-自动推导类型和输入输出
German prime minister says Ukraine will not receive "NATO style" security guarantee
How is it most convenient to open an account for stock speculation? Is it safe to open an account on your mobile phone
大数据开源项目,一站式全自动化全生命周期运维管家ChengYing(承影)走向何方?
QT establish signal slots between different classes and transfer parameters
Password recovery vulnerability of foreign public testing
5g NR system messages
[Yugong series] go teaching course 006 in July 2022 - automatic derivation of types and input and output
Qt添加资源文件,为QAction添加图标,建立信号槽函数并实现
Cve-2022-28346: Django SQL injection vulnerability