l**********n
发帖数: 226 | 1 朋友说起她的朋友在招人。大家自己看，直接联系招聘的人好啦。这里只是中介下。大
家加油。
-----
Two postdoc positions are currently available-Computational Biology, New
York City
General guideline
Prospective candidates should have a recent PhD degree in computer science,
mathematics, bioinformatics/computational biology discipline and high
motivation to pursue independent research in computational biology.
Applicants are expected to have a solid background in programming and
computational techniques, with a working knowledge of molecular b... 阅读全帖 |
|
|
c********t
发帖数: 4527 | 3 which range that you feel that you are comfortable and confident in getting
the job?
I think mine is 98k
JOB#1: Mid Level and Senior Engineer ($70k-$120k) Somerville/Davis Square
We are the world's leading music intelligence platform serving music data to
100 million music consumers a month via an API that supports hundreds of
queries a second to mobile devices, websites and applications. We are
looking for a senior engineer to help lead our team build the best music
intelligence products and gr... 阅读全帖 |
|
c*****8
发帖数: 23 | 4 有兴趣的同学请发站内信，帮忙介绍的同学如果内推成功，可平分奖金：）
Sr Software Engineer - Data Science Pipeline - Big Data, Java, Python
Job LocationsUS-IL-Chicago
Qualifications
A BS or MS in Computer Science or related field
5+ years of programming experience Java, Python, and Java Script.
3+ years of working experience with AWS services (EC2, SNS, SQS, S3, EMR,
Data Pipeline, VPC and etc.)
Development experience of Web Services (RESTful, SOAP)
Development experience with web application frameworks - Spring /Spring Boot
/Angu... 阅读全帖 |
|
w**********y
发帖数: 1691 | 5 First Semester:
Theory Of Statistics 1(STAT2631) Applied Statistical Methods 1(2131)
Applied Multivariate Analysis or Applied Advanced Time Series or Statistical
Packages
If you are interested in Finance:
Math finance I, II, III, IV@math dept. (If the instructor is Chadam)
Options@MSCF@CMU and any other available courses @MSCF
You MUST learn SAS (little SAS book; Linear Model in SAS)and R (self
included introduction) by course study or self study as soon as possible.
For other semesters:
Applied... 阅读全帖 |
|
w******4
发帖数: 488 | 6 It does not matter if you can't understand Bayes (theory), just learn how to
run one of the Bayesian softwares. Learning SAS, R or VBA alone really can
not help. BTW, I am learning Bayesian as well. Very useful tool~ |
|
e*********6
发帖数: 3453 | 7 这里的确没法用frequencist的理论来解释，毕竟中国队没在欧洲踢过很多年。只能
bayesian来做，其实Bayesian也没法做，因为P（B|A）也没法求，只能通过合理推论来
得到我的结论 |
|
O*********r
发帖数: 1835 | 8 两种概率论：Frequency versus Bayesian view. 前者若是看整个career，豆子概率大
。若是用Bayesian观点，不断用新数据修正概率，那霏霏赢面大。
有兴趣的可以读一下这个华尔街面试的传统题目：https://en.wikipedia.org/wiki/
Monty_Hall_problem |
|
w*******2
发帖数: 25 | 9 张无忌的贝叶斯世界(Bayesian world) 及其他
最近周末两天复习了93年台视版的倚天屠龙记。每年的必修课。
整部戏后半部分的主要焦点在于张无忌与赵敏之间的爱情线。张无忌对于赵敏的到底是
不是抢刀杀人的凶手的判断这件事情上，可以发现Bayesian econometrics其实可以很
好的刻画一个人对事件的observation以及他的perspective of that observation.
假设张无忌初初对于赵敏的怀疑度为P(A), 就是在此时还没有抢刀杀人事件发生。然后
，在一个荒岛上，发生了抢刀杀人的事件B，现在对于张无忌个人来说，他并不知道谁
是凶手。于是乎，他自己在主观上对赵敏是不是抢刀杀人的凶手有一个判断，即P(B/A)
, 通过剧情我们可以看出来这个概率是相当低的，因为张无忌根本不愿意相信赵敏是杀
人抢刀的凶手。他对赵敏的态度为：P(A/B) = P(B/A)*P(A)/P(B)---equation(1), 因
此，p(B/A)很低，假设杀人抢刀这件事发生与否和赵敏一点关系也没有（事实的确如此
，因为纯粹是周芷若想要光耀峨嵋派），那么P(B/A)=P( |
|
g****g
发帖数: 1828 | 10 In statistics, a priori knowledge is prior knowledge about a population,
rather than that estimated by recent observation. It is common in Bayesian
inference to make inferences conditional upon this knowledge, and the
integration of a priori knowledge is the central difference between the
Bayesian and Frequentist approach to statistics. We need not be 100% certain
about something before it can be considered a priori knowledge, but
conducting estimation conditional upon assumptions for which ther |
|
c******0
发帖数: 9 | 11 本人是经济Ph.D.，业余读了很多生物papers，而且有熟悉的本科同学在top 1 bio Ph.
D program，所以对学生物的人还是比较了解的。
首先，那破文pick on叶除了racism没有别的解释，这很多人都说的很清楚了。你要做
one-sided hypothesis testing, 肯定rejection region是挑一个区间，哪有选当中一
个discrete point的。
然而，更重要的事实是，那文章里的"performance profiling"原则犯了一个最基本而
且最常见的probability error。国人都冲着racism去了，没看到这个大漏。按照
Bayesian rule, 假设用禁药的人能50％提高performance到rejection region, 不用禁
药只有10％能提高到rejection region,然而只有5％的人用禁药，那么即使有人failed
了这个test，他用禁药的可能性也只有P=5%*50%/(5%*50%+95%*10%)=20%。也就是说，
如果单纯靠这个performance profiling t... 阅读全帖 |
|
C******n
发帖数: 9204 | 12 bayesian或者learning (but still classcial。这可以是HMM，例如filtering，但可
以把filtering可以被interpret成bayesian)都有learn-type的结果，即prior不等于
posterier。
但learning下，如果假设iid，最后一次结果仍然和前面无关，只不过人在learn。 |
|
S**U
发帖数: 7025 | 13 用Bayesian probability或是Markov chain模型解釋，就是壞事幹多了，產生不可愛結
果的機率增加。
重業就是Bayesian probability產生結果的機率大增的行為。 |
|
w********d
发帖数: 275 | 14
these are totally different concept
each model can have supervised and unsupervised case,
depends on the training data and differ is parameter estimation.
kalman filter and hmm are two kind of bayesian network,
which is called direct graphical model.
(that is ,bayesian network is direted graphical model)
kalman filter and hmm are two kind of
direct graphical model.
Also there is another kind of undirect graphical model, such as
markov random fields. |
|
N**D
发帖数: 10322 | 15 Recommend
Pattern Classification, by Duda etc. Good and comprehensive beginning book.
Learning kernel methods, by Herbrich, Good intro to kernel methods and
statistical learning theory.
Data Analysis, A Bayesian Approach. A concise and clear introduction to
Bayesian methods. |
|
s****i
发帖数: 216 | 16 我是想问个比较浅显的问题,
一般都说maximum likelihood 有overfitting的问题, 所以要加上regularization
term,
我感觉这不就是 bayesian linear regression吗?
old
Bayesian |
|
N**D
发帖数: 10322 | 17 看看Tiknov Regularization (not sure about spelling)
btw: Bayesian and some stat people 说SVM 可以这样从Bayesian 推导出来，
早干啥吃了？ |
|
N**D
发帖数: 10322 | 18 所以Bayesian 就是胡扯，但是很多funding还被所谓的Bayesian 专家控制，大家没人
敢强烈批评。
看看jmlr 的一片关于boosting 的文章，作者给了一些实验还说明现有的一些东西不对
头，在讨论的一系列文章中有一篇是大牛写的，最后原作者在总结回复中说到，类似“
我们这篇
paper 曾经先投到统计的杂志中，被莫名的锯掉，因为没有迎合大众的意思”。要等到
这些老朽们都退下去了才好。
另外， jmlr 很不错。
regularization
constraint |
|
m**s
发帖数: 346 | 19 【 以下文字转载自 Statistics 讨论区 】
发信人: meds (忘记吃药了！), 信区: Statistics
标 题: 付费找tutor
发信站: BBS 未名空间站 (Sun May 29 21:52:33 2011, 美东)
修一门bayesian methods in computer science。
主要是编程实现一些统计上面的算法。
开始还好，最近的一个lab，让我很头大。也许给我很多时间，我能想出来，但是会太
迟了。明天就要交了，虽然是memorial day。
不是想找人帮忙做，只需要懂的人看看，点拨一下。为了表示对您的时间的尊重，我愿
意$10/hr 补偿您的时间。钱肯定不多，您如果嫌不够，可以再商量。
下面是这个lab的大概意思：
implement MCMC using Metropolis (or Metropolis-Hastings).lab给了4个bayesian
networks，要用我们写的程序针对某个query，做sampling，然后plot mixing, prior,
and posterior. 其实我自己感觉这个lab有些地... 阅读全帖 |
|
g**********t
发帖数: 475 | 20 其实最大的缺点是MCMC太慢了。我们实验室有一个人下了一个MCMC软件算一个比较大的
数据集，已经算了三个多月还没有收敛，汗~~
不过Bayesian逻辑上看起来很爽，我已经把frequentist看作是Bayesian的一个特例了。 |
|
d*****u
发帖数: 17243 | 21 我感觉很久很久以前，data mining和machine learning几乎是同义词
后来machine learning的主流走了统计的路子，
而data mining保留原来的味道
还有就是怎么把一些现实情况转化成可计算的data
也就是representation的问题，
可能machine learning的人做得要多些，但是也不尽然
还有,machine learning的人用bayesian statistics的比较多
而在统计学领域里，bayesian和frequentist是有斗争的 |
|
K****n
发帖数: 5970 | 22 Bayesian的问题本版牛人们好像也早就批判过了，凡是支持Bayesian school的，都是"
error bar
people" |
|
s******s
发帖数: 13035 | 23 补充一下为什么选了one-tail就不能换了。其实，前面bigsail已经说了，
hypothesis test model是理论作出的，不是根据实验数据作出的。如果你
根据理论选了单尾model，就算实验数据不符合，也不能justify你换model。
这也就是我前面有一片文章里面说的统计哲学问题，同样也是我前面说过
不喜欢经典统计，而喜欢bayesian的原因。bayesian的好处就是简单明了，
而且把这些先前的经验（包括根据数据得出的结果）都整合在model里面，
没有很多对新手来说事实而非容易混淆的东西。
英国著名政治家，维多利亚时期的首相本杰明·迪斯雷利说过“ “There
are three types of liars: liars, damned liars and statisticians.”
我听过的第一门统计课，开场白就是“统计学家要自律”。拿到data以后，
其实不根据这些“统计哲学”来处理，其实想拿到什么结论就能得出什么
结论。对生物学家来说，如果one-tail不对，保证99.9%的人马上变成two-
tail或者换成反方向one-tail，这个... 阅读全帖 |
|
l**********1
发帖数: 5204 | 24 生物统计的入门教材之一 哈 of course english version:
http://www.roma1.infn.it/~dagos/rpp/
Introduction
Uncertainty and probability
Rules of probability
Probability of simple propositions
Probability of complete classes
Probability rules for uncertain variables
Bayesian inference for simple problems
Background information
Bayes' theorem
Inference for simple hypotheses
Inferring numerical values of physics quantities -- General ideas and basic
examples
Bayesian inference on uncertain variables and posterior ... 阅读全帖 |
|
l**********1
发帖数: 5204 | 25 不要只看到Bayesian+Markov Chain Monte Carlo (MCMC) 用于NGS sensitivity
analysis
新的非 Bayesian+MCMC的 Fisher Information Matrix (FIM) sensitivity
analysis 已有英国 帝国理工大的论文了吧 吧
Reference:
i,
Bioinformatics. (2012) 28:731-3.
web link:
HTTP://www.ncbi.nlm.nih.gov/pubmed/22378710
PDF file:
HTTP: //www.theosysbio.bio.ic.ac.uk/wp-content/uploads/2011/04/manual1.pdf
就是由减少Stochastic simulation process 中的参数 数量的 视点入手的
关联论文
ii,
J R Soc Interface. (2008) 5 Suppl 1:S59-69.
HTTP: //www.ncbi.nlm.nih.gov/pubmed/1848... 阅读全帖 |
|
l**********1
发帖数: 5204 | 26 名校 新人PhD 的 有关 关键词的 博士论文 最近三年到五年的 至少100篇
其全部PDF 文档或hard paper 都参考一下的话 可以看出点 如何建筑这幢楼的 端倪了
吧 (possible)
比如
citation from UBC 2009 PhD dissertation: "Particle Markov Chain Monte Carlo"
by
Roman Holenstein
full text pdf link:
HTTP: //www.cs.ubc.ca/grads/resources/thesis/May09/Holenstein_Roman.pdf
Monte Carlo methods have become the standard tool to solve many problems in
statistics and scientific computing. Examples are abound, and include
instances in
Bayesian statistics (posterior estimat... 阅读全帖 |
|
l**********1
发帖数: 5204 | 27 正在从 旧约 向 新约 过渡吧
pls refer:
i,
S. Geman and D. Geman (1984)
Stochastic relaxation, Gibbs distributions, and the Bayesian restoration
of images.
IEEE Trans. Pattern Anal. Mach. Intell., Vol. 6, 721–741.
PDF web link:
HTTP : //www.stat.cmu.edu/~acthomas/724/Geman.pdf
ii,
Anisotropic Diffusion in Image Processing. (2008)
by Joachim Weickert
PDF web link:
HTTP ://www.lpi.tel.uva.es/muitic/pim/docus/anisotropic_diffusion.pdf
iii,
Glaus P et al. (2012)
Identifying differentially expressed transcripts ... 阅读全帖 |
|
H*g
发帖数: 2333 | 28 I am not studying evolutionary biology neither. Just my 2cents. I guess I
would forget the neighbor-joining method.
Assuming that you have a good alignment with the most biological sense (
functional domain well aligned), good bootstrap value for ML and good
convergence for Bayesian Inference, regarding of the trees that are
generated from those different methods, you may have to distinguish them by
yourself by figuring out which one makes the most biological sense.
Without a thorough look at th... 阅读全帖 |
|
l**********1
发帖数: 5204 | 29 just occasionally met this post,
and all sofa floor was/were only copy and paste from one proposal which sent
on last week what is about lncRNA and
machine learning...(ML) stuff n.b. not make love :-)
if juzbox (Drug Resistance) also has interest,
pls check,
video tube of online lecture > Bayesian and ML,
http://machine-learning-course.joachims.org/
or one TT AP posiiton its description,
http://bayesian.org/forums/jobs/6949 |
|
发帖数: 1 | 30 如果是转一般的马工，实在是没多大意思，也太浪费你男友的数学底子。要转也要去做
科研，比如说做Bayesian machine learning。据我所知，绝大多数Bayesian
inference方法（如各种MCMC方法和variational inference）都是那些做机器学习/统
计的大牛从统计物理学家那里抄过来的，你男友背景正合适， |
|
J**Y
发帖数: 34 | 31 I will use Bayesian approach in my thesis. Tony Lancaster just came to
our department and did a presentation "Orthogonal parameters and panel data".
I don't know his work in Bayesian. I just know some big names such as
Zellner, Geweke. In addition, a professor in our department-Dale Poirier,
who just just transfered to UCI from U. Toronto, is very famous
in this field. |
|
t******s
发帖数: 39 | 32 To idioteque:
握手。同是宏观经济学生。
用Bayesian方法估计新凯恩斯主义DSGE模型时，对于prior distribution的设定和weak
identiciation的问题的确有争论，但这不能否定Bayesian估计方法，它在理论上是站
得住脚的。在我们所关心的dimension上，prior standard error自然设定大些。如果
tight prior对于posterior估计结果和marginal data density影响很大，那这样得到
的结果自然不robust,负责任的研究不会这样做，但tight priors在有些时候是可以接
受的。0.0001的prior standard error不常见。
“我的观点是，宏观经济学研究者（包括我自己）所了解和掌握的宏观经济规律不多，
现在的方法存在很多问题，还是承认我们的无知更好。” |
|
w****a
发帖数: 155 | 33 对 game theory 中的两个概念比较疑惑:
Bayesian stackelberg game和 stackelberg game 的关系是什末？
Bayesian stackelberg game 可以看作 singaling game 的一个特例吗？ |
|
h****i
发帖数: 1674 | 34 hierarchical model in bayesian
你不用手写啊，随便看一会儿bayesian相关note你就自己能写code了 |
|
D******n
发帖数: 2965 | 35 顺便再多请教一个问题：对于Bayesian Game, 大家都假设type的分布是common
knowledge, 我记得game 课上讲过，好像早期对于这个假设有很多争议，但后来有一篇
文章出来，说在很简单的一个Bayesian game的setting，如果不假设common knowledge
, 那不管什么结果都可以是均衡，然后关于 common knowledge假设的争议就统统消失
了。请问你（或其他版上朋友）知道这个文章的出处吗？ |
|
B****n
发帖数: 11290 | 36 一個簡單的方法就是先主觀的定它的難度(依據需要思考的步數 分支 有沒有需要"妙手
"等)
然後呢把它放線上由使用者來玩 按照答對的程度 來adjust
這就屬於Bayesian model 先有一個prior distribution(主觀難度) 再由使用者答題狀
況來update你的distribution 得到所謂的posterior distribution 很多線上排序的東
西就是用Bayesian model來弄得 |
|
i**f
发帖数: 1195 | 37 hehe, the empirical bayesian individual estimates are not real bayesian
estimates...
from |
|
c*********g
发帖数: 154 | 38 这个是典型的Bayesian Inference啊。在用Bayesian的时候，最重要的是分清什么是先
验和后验，其实说到底，就是分清楚原因（或者叫状态）和结果（或者叫观测）。所以
一般的做法都是这样的：
1.列出所有可能的原因（状态）及其先验。对于这两道题，状态集合是{BB,BG,GB,GG}
，并且P(BB)=P(BG)=P(GB)=P(GG)=1/4,
2.列出从各原因（状态）到结果（观测）的条件概率分布（离散的或连续的）。
对于Problem 1: P(告诉别人我有一个女孩 | BB) = 0,
P(告诉别人我有一个女孩 | BG) = 1,
P(告诉别人我有一个女孩 | GB) = 1,
P(告诉别人我有一个女孩 | GG) = 1,
对于Problem 2：P(别人看到我有一个女孩 | BB) = 0,
P(别人看到我有一个女孩 | BG) = 1/2,
|
|
T*****w
发帖数: 802 | 39 【 以下文字转载自 JobMarket 讨论区 】
发信人: watercool (yy), 信区: JobMarket
标 题: 贴几个关于quant的职位
发信站: BBS 未名空间站 (Mon Oct 25 21:37:29 2010, 美东)
这些职位都在nyc地区的major financial institutes
细节请联系f*****************[email protected]
=====
Quantitative Analyst
Main Duties :
• As a quantitative analyst you will be responsible for
independent verification and validation of the models and pricing
functions
• Supporting business growth by independently verifying
implementation of new products within the front-offi... 阅读全帖 |
|
c*******e
发帖数: 150 | 40 看看天花乱坠忽悠到头了，其实挺空虚的，基本都是大陆技术 factor models -> mean
-variance portfolio construction with a power-functional form of TC.
最后一段 "Second-Order Cone Programming" 是告诉观众们 他们的 portfolio
construction process 除了一个 TE constraint，其它的都是线性constraints，而且
the TC term 应该是 有理数次方 （e.g. trades.^2 or trades.^(3/2) ）。 只有
有一句话“techniques of Bayesian return forecasting” 偶没有完全看懂，不知道
是说在return forecasts过程中用了 Black-Litterman shrinkage，还是说用了dummy
observations 这些经典的Bayeisan 技术，还是有更高级的Bayesian 技术。 抛砖引玉
，希望有大侠出来指点一二 :) |
|
s*******i
发帖数: 546 | 41 如果想做quant的话
Statistical Inference
Econometrics
Regression and Statistical Computing
Time Series Analysis
BAYESIAN DATA ANALYSIS
Bayesian Statistics
Multivariate Analysis
Extreme Value Theory
Data Mining and Statistical Learning
Stochastic Differential Equations
外加个这门课:
Statistical Simulation (random number generators; Monte Carlo methods,
resampling methods, Markov Chain Monte Carlo, importance sampling and
simulation based estimation for stochastic processes) |
|
d**********n
发帖数: 16 | 42 其实很多CS PHD的数学是非常NB的，尤其Graphics, AI, Theory,只不过Stochastic在
CS自己的领域里应用较少，所以学的人不多，好像解微分方程，通常也就Graphics里模
拟流体(电影里动画做成的海啸)的时候用用，其它真用的不多。
现在AI走的是Data-Driven的路线,不是Model-Driven。不是先想个model,再用data验证
,而是根本没model, 靠黑盒(svm, adaboost, decision-tree,bayesian..) 再加海量
data, train出来的。 70,80年代,搞人脸识别,文章分类的人也用过微分方程,和一些物
理研究的手段,都失败了,90年代bayesian学派兴起之后,逐渐开始了Data-Driven的策略
,现在的kinect, 照相机的人脸detection, google的各种关于文章的分类,翻译,都是这
些东西的产物。
不过感觉现在很多HF也开始Data-Driven的策略了。 |
|
l********a
发帖数: 126 | 43 另外，vol trading（trade option或者variance swap）里用的vol forecast本质上也
是Bayesian inference吧？
因为除了market-implied和historical data似乎没有其他的information source来预
测vol。而这两个information source，据我所知没有什么单一的model能把它们统一到
一起。所以唯一的把两个统一到一起到的方法就是Bayesian inference了（或者更
sophisticated的Machine learning方法）。
我不做vol trading，不知道以上猜测对不对。
) |
|
k***g
发帖数: 7244 | 44 We use Bayesian games for systematic trading and market making. (we don't
derive, say, perfect Bayesian equilibrium formally but we use the intuitions
)
I think game theory is more useful for a relatively small set of players
such as primary market makers. When there are many players, though we can
assume players are homogeneous in such a way that they can be represented by
a few agents, it is not very useful practicality... |
|
I**a
发帖数: 26 | 45 Job posting for a friend, who works for a emerging hedge fund in New York
City -
Her team is expanding fast and is actively seeking quantitative researchers
and data scientists who will be responsible for independently conducting
quantitative finance research with a focus on data-driven statistical and
predictive models.
Ph.D. or postdoc in a quantitative field is preferred.
Experience processing and analyzing large datasets is a plus.
Experience in signal processing and time series analysis is ... 阅读全帖 |
|
m*****e
发帖数: 207 | 46
I don't think so. Why are they (hehe, I'm Bayesian) called Frequentists?
The answer is: they think probability is the limit of frequencies, as
the number of repeated experiments approaches infinity. They believe
that probability is an objective matter.
However, Bayesians believe everybody can have his/her own view on a
probability that an event happens, i.e., probability is subjective, not
objective. For example, even with a same coin, people may have different
views on the probability of |
|
m*****e
发帖数: 207 | 47
你混淆了鲜艳概率和已知概率
这么说吧，如果是频率派，他对p的估计一般是用频率估计概率
比如他经过1万次实验发现有5000次正面，他就估计p=1/2
如果是贝叶斯派，他也可以“先验”的认为p=1/2: 这实际上是
p的一个单点分布：Pr(p=1/2)=1. 尽管Bayesians一般都不选这
个先验分布。
原题里实际上是假设p=1/2，没有提p到底是固定的数呢还是随机变量
所以可以不考虑到底是Frequentist和Bayesian |
|
|
f******a
发帖数: 9 | 49 The Variational Bayesian EM Algorithm for Incomplete Data: with Application to
Scoring Graphical Model Structures (2003)
Matthew J. Beal, Zoubin Ghahramani
Bayesian Statistics 7
pdf link:
http://citeseer.nj.nec.com/beal03variational.html
而
作 |
|
