Actuarial Outpost
 
Go Back   Actuarial Outpost > Actuarial Discussion Forum > Software & Technology
FlashChat Actuarial Discussion Preliminary Exams CAS/SOA Exams Cyberchat Around the World Suggestions

DW Simpson
Actuarial Jobs

Visit our site for the most up to date jobs for actuaries.

Actuarial Salary Surveys
Property & Casualty, Health, Life, Pension and Non-Tradtional Jobs.

Actuarial Meeting Schedule
Browse this year's meetings and which recruiters will attend.

Contact DW Simpson
Have a question?
Let's talk.
You'll be glad you did.


Reply
 
Thread Tools Display Modes
  #1  
Old 12-10-2017, 03:21 PM
clarinetist clarinetist is offline
Member
Non-Actuary
 
Join Date: Aug 2011
Studying for Rcpp, Git
Posts: 6,788
Default Machine Learning and Pattern Recognition Thread - starting 12/15/17

This will be a very theory-driven study. If you're just interested in running code and none of the theory, I would recommend you look elsewhere. There are many books and resources that serve this purpose.

I'm taking a grad-level statistical machine learning course next semester and am trying to learn this material ahead of time. My goal is mainly to learn the theory, with implementing the algorithms in R (maybe some Python) on the side.

Here are the books I'll be using, roughly from highest to lowest emphasis (primary texts in bold):

Machine Learning: A Bayesian and Optimization Perspective (Theodoridis)
Pattern Recognition (Theodoridis)

Pattern Recognition and Neural Networks (Ripley)
Pattern Recognition and Machine Learning (Bishop)
Foundations of Machine Learning (Mohri et al.)
Understanding Machine Learning (Shalev-Shwartz, Ben-David)
Hands-On Machine Learning with Scikit-Learn and Tensorflow (Geron)
Machine Learning (Flach)
Make Your Own Neural Network (Rashid)
Modern Multivariate Statistical Techniques (Izenman)
Principles and Theory for Data Mining and Machine Learning (Clarke et al.)
Elements of Statistical Learning (Hastie et al.)
Machine Learning (Murphy)
Learning from Data (Abu-Mostafa et al.)
Introduction to Statistical Learning (James et al.)
Practical Data Science with R (Zumel, Mount)
Python Machine Learning (Raschka)
Applied Predictive Modeling (Johnson, Kuhn)
Bayesian Methods for Data Analysis (Carlin, Louis)
Generalized Additive Models (Wood)
Matrix Computations 4th ed. (Golub, Van Loan)
Convex Optimization (Boyd, Vandenberghe)

I will be starting with the Machine Learning text by Theodoridis, and then proceed to the Pattern Recognition text when finished with the ML text. I will also post blog posts explaining the concepts as I learn them, as well as code in primarily R (maybe Python) where appropriate.

This will be the main thread for discussion on these concepts. I will not have a predetermined schedule, but will try to cover concepts as quickly as possible. I'm going to assume that you are proficient with probability (exam-P level is fine), linear algebra, and mathematical statistics at an undergraduate level and will not spend any time reviewing this material. However, if you have any questions on the material, please feel free to ask in this thread.
__________________
If you want to add me on LinkedIn, PM me.

Why I hate Microsoft Access.

Studying/Reading: Machine Learning and Pattern Recognition

Last edited by clarinetist; 12-10-2017 at 03:28 PM..
Reply With Quote
  #2  
Old 12-16-2017, 03:07 PM
clarinetist clarinetist is offline
Member
Non-Actuary
 
Join Date: Aug 2011
Studying for Rcpp, Git
Posts: 6,788
Default

Section 2.2.4, on complex random variables: http://ymmathstat.blogspot.com/2017/...variables.html

Sections 2.4 and prior should be mostly review (or slight extensions) of exam-P material.
__________________
If you want to add me on LinkedIn, PM me.

Why I hate Microsoft Access.

Studying/Reading: Machine Learning and Pattern Recognition
Reply With Quote
  #3  
Old 12-18-2017, 02:39 PM
DiscreteAndDiscreet DiscreteAndDiscreet is offline
Member
AAA
 
Join Date: May 2016
Posts: 385
Default

I’d say that I think most of this syllabus ties back to computational complexity with the understanding that, subject to some feasibility conditions, generalized inference is an NP optimization problem (and I’ll speculate that it’s NP complete as well).

A complex random variable shouldn’t be much more problematic than a vector valued random variable. I’ve had to work with complex scalars in a quantum computing course and the way it worked was that the formalism had too many degrees of freedom to represent the system but you basically either add a constraint to compensate or you identify what values of the variables need to be treated as physically indistinguishable.

Complex scalars give linear algebra a little more oomph in terms of being able to treat more phenomena as purely linear. You can, of course, represent complex numbers as matrices and linear transformations on them can correspondingly be block matrices but making them scalars emphasizes that the same math can be used for the full range of phenomena.
Reply With Quote
  #4  
Old 12-18-2017, 03:11 PM
BassFreq's Avatar
BassFreq BassFreq is offline
Member
CAS
 
Join Date: Jun 2003
Location: Chicago
Studying for all eternity
Favorite beer: Duff
Posts: 1,691
Blog Entries: 2
Default

__________________
If at first you don't succeed, you have one data point.
Res ipsa loquitur, sed quid in infernos dicet?
Reply With Quote
  #5  
Old 12-18-2017, 11:21 PM
ultrafilter ultrafilter is offline
Member
 
Join Date: Nov 2006
Posts: 373
Default

I've never seen complex random variables in the context of machine learning. What's the connection here?
Reply With Quote
  #6  
Old 12-19-2017, 07:49 AM
clarinetist clarinetist is offline
Member
Non-Actuary
 
Join Date: Aug 2011
Studying for Rcpp, Git
Posts: 6,788
Default

Quote:
Originally Posted by ultrafilter View Post
I've never seen complex random variables in the context of machine learning. What's the connection here?
Stay tuned, I'm still trying to figure that out. It appears that they're being referred to throughout the text.
__________________
If you want to add me on LinkedIn, PM me.

Why I hate Microsoft Access.

Studying/Reading: Machine Learning and Pattern Recognition
Reply With Quote
  #7  
Old 12-19-2017, 06:20 PM
DiscreteAndDiscreet DiscreteAndDiscreet is offline
Member
AAA
 
Join Date: May 2016
Posts: 385
Default

FFT.

Last edited by DiscreteAndDiscreet; 12-19-2017 at 06:30 PM..
Reply With Quote
  #8  
Old 12-19-2017, 07:25 PM
DiscreteAndDiscreet DiscreteAndDiscreet is offline
Member
AAA
 
Join Date: May 2016
Posts: 385
Default

The largest absolute value component of an FFT run on a music sample gives you an estimate of the beats per minute of the music. If you bootstrap the sample in a manner that preserves the frequencies that you're interested in, you can get a distribution of spectra that should contain more inferred features.

Or, let's say you take a greyscale image of text and generate random line segments and run an FFT on the greyscale values that intersect with the line segment, you should get a distribution of spectra that relates to the presence of features defined by the relative distance of dark/light pixels that would not be sensitive to orientation of the text. If you give the line segments a distribution of measures that gives them coverage of adjustments for viewing text from different angles, it could reduce sensitivity to different viewing angles.

I'm not sure how efficient either of these concepts would be, but they seem like they should pick up some feature inferences.
Reply With Quote
  #9  
Old 12-20-2017, 10:05 AM
DiscreteAndDiscreet DiscreteAndDiscreet is offline
Member
AAA
 
Join Date: May 2016
Posts: 385
Default

A more efficient version if that second concept is outlined here:

http://www-sigproc.eng.cam.ac.uk/fos...sApps_UDRC.pdf

Actually some of this ties back to observations of animal vision systems. Presence of very simple features that can be read with a kernel density estimator or wavelet correspond with the firing of specific neurons in animal studies.
Reply With Quote
  #10  
Old 12-20-2017, 11:13 AM
ultrafilter ultrafilter is offline
Member
 
Join Date: Nov 2006
Posts: 373
Default

Fourier methods use complex numbers, but to the best of my knowledge they don't specifically rely on complex random variables.

I found a CAS note that describes complex random variables and some potential actuarial applications. I don't know how mainstream those ideas are or will be, but odds are good that someone will find it useful.
Reply With Quote
Reply

Thread Tools
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off


All times are GMT -4. The time now is 09:31 PM.


Powered by vBulletin®
Copyright ©2000 - 2018, Jelsoft Enterprises Ltd.
*PLEASE NOTE: Posts are not checked for accuracy, and do not
represent the views of the Actuarial Outpost or its sponsors.
Page generated in 0.14339 seconds with 9 queries