p********n
发帖数: 273 | 1 Hi experts,
I got the fitting result of the gmdistribution object. Now I want to know
the exact expression of the 2D guassian function. From Wikipedia, I know the
expression is
f(x,y) = A*exp(-(a*(x-x0)^2 + 2*b*(x-x0)(y-y0) + c*(y-y0)^2))
where A is the hight of the peak and (x0,y0) is the center of the blob.
As far as I know,
x0 = obj.mu(1)
y0 = obj.mu(2)
My question is how to get coefficients a b c from obj.Sigma?
I supposed obj.Sigma = [a b
b c]
but that was n |
|
l******n
发帖数: 4 | 2 or the mean and variance of the inverse of a Guassian RV?
Thanks |
|
n*******d
发帖数: 650 | 3 Guassian filter can be generated in matlab by one function. I forgot its name.
try guassian or something like it.
In fact, any vector can be a filter such as :[.1,.3,.9,.3,.1] and you have to
normalize it, of course. for low pass filters, big elements in the middle,
small ones on both sides. for high pass, reverse.
tell more why your @ point is the answer. there must be a prior knowledge to
find the answer. you can't just give some data and segment it, i think
neighbors |
|
r***e
发帖数: 29 | 4 how about assuming the distribution? The sum of a Guassian is a T-
distribution. |
|
z**w
发帖数: 69 | 5 very simple:
1. duplicate your image into a new layer;
2. Guassian blur it to some extent;
3. adjust the opacity of the blurred layer and get ur desired effect.
BTW, can I take a look at the wedding pics u took recently? |
|
|
s*****e
发帖数: 21415 | 7 我具体不是做语音的，等我开始学习的时候都已经是mixture guassian HMM 一统江湖
了，呵呵。
所以早期的工作了解的不多。。。 |
|
s******e
发帖数: 285 | 8 看Carl的这本书，讲的很仔细。
http://www.gaussianprocess.org/gpml/
简单的来说就是个stochastic process, assuming a
GP prior.
一般的算法是假设input x是normal distribution,
GP是假设function space f(x)是normal distribution.
对于regression来说，因为y的值是连续的，所以
likelihood function p(y|f(x))也可以表示为Gaussian
的形式，这样一来posterior p(y|x) = \int p(y|f(x)) p(f|x)
就也是guassian的了，due to conjugation.
Classification的情况复杂一些，需要一个 multi-logic
function来转换p(y|f(x)).
Christ Bishop的书也写得很浅显易懂。 |
|
D***r
发帖数: 7511 | 9 Contemporary machine learning has nothing to do with "algorithm" in a
traditional sense.
I mean, in an algorithm class, you will not learn the fundamentals of the
mainstream machine learning techniques: linear regression, kernel methods,
Guassian process......
Typically, someone with a master's degree does not know enough to be a
decent engineer of this expertise. But there are always exceptions. |
|
t*******f
发帖数: 2634 | 10 Hi Could someone please help me to find where "gauss.c" can be found?
I am interested in the original author who wrote it.
It might be from "numerical recipe in C"
But I do not have access to it.
The following is part the comments from the gauss.c file:
/***************************************************************************
***/
/* DESCRIPTION: This function perform the Guassian Elimination algorithm.
INPUTS:
array -- the input Martrix stored in the vectore
numvar - |
|
e*****r
发帖数: 379 | 11 Duty cycle: see wiki, but here are some more hints...
NL, parent ion functions--Quadruples (filter) are used. For every scan, such
as 1 sec for one scan with 101-200 m/z (100 m/z span): for every m/z unit (
actually a few more data points, since mass profile is mostly near to
guassian), it gets only 1/100 second time. In another word, considering one
m/z unit: only 1/100 second was used for accumulating a given m/z, and 99%
of the scan time was "wasted" for other ions within this mass range with |
|
b******h
发帖数: 2715 | 12 刚发现上了十大。 鞠躬谢谢大家的关心。我一定会据理力争的。
既然是据理，借这里的人气问一下有没有Guassian里计算excited states 的模板？ 我
看了很多forum，可是还是不能写好syntext，我目前的方法是ub3lyp/6-31g(d,p) cis.
我的有机分子是个自由基，希望算出产生这个自由基需要的光波波长。
再次谢谢大家。 |
|
b******h
发帖数: 2715 | 13 借这里的人气问一下有没有Guassian里计算excited states 的模板？ 我
看了很多forum，可是还是不能写好syntext，我目前的方法是ub3lyp/6-31g(d,p) cis.
我的有机分子是个自由基，希望算出产生这个自由基需要的光波波长。
如果有大牛可以帮到我请电邮联系我也行，可能需要看具体的chemdraw分子。
再谢过 |
|
S****8
发帖数: 401 | 14 现在不少lab用Guassian啥的simulate一些Molecular，做Quantum cellular automata
这种东西，求问大拿这算promising的东西？还是坑？ |
|
c*******e
发帖数: 8624 | 15 guassian quad is an algorithm |
|
h***o
发帖数: 539 | 16 need more detail about F and e
if F and e have good features, X is a Guassian distribution, its mean and
variance can be obtained by solving two ODEs.
Otherwise, Monte Carlo |
|
h*******o
发帖数: 778 | 17 in fact one colleague in my group is designing a hardware random generator
for generating Guassian noise. You can google there are some paper
describing
using a complicate square/cosine functions to generate random number which
has avery very large reapting cycle... |
|
|
l******n
发帖数: 4 | 19 E[1/x] where x is Gaussan doesn't exist? |
|
|
v****k
发帖数: 229 | 21 NO.
suppose g(x) is a function, and x follows some distribution f(x), then
E[g(x)] exists if the integral \int{g(x)f(x)}dx absolutely converges,
i.e. \int{|g(x)|f(x)dx}
for your example, suppose x follows N(\mu,\sigma^2) you can check the
integral \int{(1/|x|)f(x)} diverges. |
|
l******e
发帖数: 470 | 22 \int_{-\infty}^{+\infty} Q_1(x)Q_2(x)Q_3(x) dx
Q_1,Q_2,Q_3是3个不同的guassian的culmulative distribution
多谢。 |
|
H***a
发帖数: 735 | 23 请问是"guassian的culmulative distribution"? <-这个什么意思？
还是pdf - probability density function?
如果是pdf, 我觉得3个Q一定能整理成一个新的Gaussian pdf - QQ(x) * exp(const.)
\int_{-\infty}^{+\infty} QQ(x)*dx = 1
结果就剩下那堆const, exp(const)的形式。你说呢？ |
|
l******e
发帖数: 470 | 24 aX
a=0 w.p 0.5, =1 w.p. 0.5
X = N(0,1)
就不是 guassian
gaussian不可能有0.5的概率＝0 |
|
z****n
发帖数: 17 | 25 E[X-Y| 2X-Y=c]
c=X +X-Y, therefore X-Y=c-X
E[X-Y|c]=E[c-X|c] =E[c-X]=c
c =2X-Y therefore is a guassian.. N(0, VAR(2X-Y))
Easy to compute VAR(2X-Y)=4\sigma_x^2 + \sigma_y^2 + 4\sigma_x \sigma_y \rho_{x,y} refer to previous
post..
In the end E[X-Y| 2X-Y] follows N(0, VAR(2X-Y)) |
|
T*******t
发帖数: 9274 | 26 区别应该不大吧。反正x_t是个guassian
分？ |
|
w**********y
发帖数: 1691 | 27 我是真不明白....
不是应该 (w/s1)*X1+((1-w)/s2))*X2么?
variance是 (w/s1)^2*0.2^2+((1-w)/s2))^2*0.3^2+(w/s1)((1-w)/s2))*0.2*0.3
你这里面有s1和s2啊..除非你给s1=s2.那结果就是6/7..
第一问是0,E(sin(W_t))=\int sin(x) f(x/sqrt(t))dx, where f(x) is guassian
density. f(x),sin(x)都是symmetrical的,所以积分是0.
第二问对的-不是martingale.
我猜想,他先问你第一问,就是为了迷惑你让你第二问回答是martingale吧?所以觉得这
个题稍微有意
思一些.. |
|
f**h
发帖数: 34 | 28 先算出序列x_j-x_{j-1}, 然后应该可以先test出sigma不是常数的一个，然后再fit
guassian区分出1）和3）。不知道对不对。 |
|
g******e
发帖数: 352 | 29 谢谢，wiki的图还是精确多了
看起来platykurtic的pdf在sigma 到 2 sigma这段比gussian fat,
但再远到真正tail的地方就又比guassian低了. |
|
y**t
发帖数: 50 | 30 chebyshev quadrature is the guassian quardrature over the
integral [-1,1] with weighting functionW(x)=(1-x^2)^{-1/2}
The abscissas for quardrature n are given by the roots
of the chebyshev polynomial of the 1st kind T_n(x)
and there are formulas about the wights too
some example
n abscissas weights
2 +/-0.7071 1.5708
3 0 1.0472
+/-0.8660 1.0472
4...... |
|
g***u
发帖数: 22 | 31 Use a blade and a 1-D precise positioner and a power metter. Put the power
meter behand the blade which is mounted on the positioner. Move the postioner,
measure the the Power as a function of position, fitted with cummulative
guassian funcion
it |
|
n******0
发帖数: 298 | 32 本人对此不太熟，想请教有经验的大牛。如果算5-10 dimensional 的数值积分（比如
都是normal distribution)，有什么好的方法和现成的软件？我的理解是传统的方法，
比如Guassian quadrature,在这么多维的情况下，太繁琐，计算速度也会很慢。我没试
过，不知道说的对不对？还有一个我知道的方法是monte carlo integration，不知道
精确性如何？ |
|
l******n
发帖数: 9344 | 33 问个问题哈
n个source,每个source sample,因为观测限制，Ti,i=1,...,n以下的都没有记录。现在
假设每个都是guassian,要做mixture model.那怎么做？
谢谢 |
|
l*******s
发帖数: 26303 | 34 这周比较闲，俺也来滥竽充数贴一个，基本思路是这样的：
1) 树叶和树丛上的积雪。这个最精确的办法是手绘，最好手边有几张可以参考的照片，看看树叶和
树丛积雪的有什么物理特点，用白色和浅灰/蓝色绘出积雪的明暗立体感来。后面lemma和冬盖都是
用的这个办法。俺比较没有耐心就偷懒了一把，用的是select -> color range，选择树叶上的
高光部分，多试几次，一直选到自己满意的选区为止，然后新建层用白色填充，这个就是积
雪“层”了。基本上画面上树叶、石头、船篷的高光部分都能积上“雪”，选取的时候不可避免会选到很
多无关的地方如水面和房子的白墙等，这可以用不同形状和笔触的橡皮擦擦掉。觉得应该积雪深点的
地方如船篷，可用不规则的毛笔画点上去。不过这个偷懒的办法没法体现积雪的厚度和立体感 -_-。
2)天空中的飘雪。新建层用filter->Gaussian noise，再稍微gaussian blur一下，然后
Image->Threshold，调整一下threshold的level，直到雪花的密度自己满意为止。这时候雪花
很小，可考虑拉大几倍。再来一次guassian blur和mot... 阅读全帖 |
|
