关于Noise and Error主题的一些小知识

本文探讨了噪声对VCbound的影响,并得出结论:噪声不会对其产生本质影响。此外,还讨论了机器学习中误差评估函数err与实际使用的err'的区别及其原因。

摘要生成于 C知道 ,由 DeepSeek-R1 满血版支持, 前往体验 >

(一)Noise会不会对VC bound产生影响?

此笔记源于台湾大学林轩田老师《机器学习基石》《机器学习技法

 

答案是不会。

当信号中加入了Noise,其实对我们之前学过的内容不产生任何本质上的影响。

之前《机器学习真的起作用吗?》中的分析是基于:training dataset D的来源是:(x,f(x))。其中x从服从某一概率分布P。

现在noise对数据产生了什么影响?之前,f(x)是确定的。现在f(x)不确定了。因为Noise的存在会影响f(x)的值。此时y更像是服从某一概率分布。但是抽样的一瞬间,y是确定的值。我们要估计的就是抽样这一瞬间,整体中的x和y。

image

如上,根本不影响我们的理论分析。

 

                                                        (二)err与err’

根据之前我们的学习,我们知道,Learning Algorithm A的一个重要的原则是:使得Ein很小(即选择的g在inside D上犯的错误很小)。

g在D上犯得错误记作err。

同时我们知道,错误的评估函数也是A的重要组成部分,记作err’。

理论上来说:err’=err.

但是在实际中很难实现,我们甚至无法准确地描述err是什么样子的。又或者err非常不利于计算。

通常情况下,我们使用不同于err的err’作为A的key part。

image

image

上面这个流程图是一个机器学习的基本流程:其中我们使用err’进行学习,使用真正的err进行评估。

转载于:https://www.cnblogs.com/wangyanphp/p/5439021.html

A = imread('krabi1.bmp','BMP'); %read in the image A_shrunk = imresize(A,0.2); % we’ll reduce the resolution, as otherwise the file size is too large imshow(A_shrunk) % displays the shrunken image Bs = reshape(A_shrunk,[388*518*3,1,1]); % resizes this image from a pixel array of three colours to a one-dimensional data stream C = de2bi(double(Bs)); % converts these values to binary representation % You can then resize this array to a linear, one-dimensional array. % this data stream C is then what you can communicate over your channel. % recover the image from the binary sequence BS_rec = uint8(bi2de(C)); % convert bits to pixel values. A_rec = reshape(BS_rec,[388,518,3]); % reshape back to a coloured pixel array. imshow(A_rec) % display the recovered image. Explore the pskmod and pskdemod commands in MATLAB. Use these to modulate and demodulate the binary sequence into a binary PSK signal. (Hint: the command should be something like D = pskmod(C,2). Display a part of this signal, and explain carefully how we have represented this signal in MATLAB, and how it differs from the analysis we performed in Part A. Transmit this signal through an AWGN channel using the awgn() function for a range of signal to noise ratios. Demodulate and comment on the effect of the noise on the image. A suitable range of signal to noise ratios is –0.5dB to 5 dB. The bit error rate refers to the fraction of bits that are received in error. This can also be interpreted as the probability that a bit is received in error. Calculate the bit error rate (BER) for the signal to noise ratios you used above. Comment on the correlation between BER and perceptual image quality. Now apply channel coding to the transmitted data sequence. Use encode(msg,7,4,’hamming’), encode(msg,15,7,’bch’), and convenc, following the example listed in the help window (recall, in MATLAB type “help convenc” to obtain the help information for that command. For each of these codes, study how the code can reduce the BER and improve image quality for each SNR you studied above.
06-10
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值