Options

Information theory and the El Nino project

I learned some truly amazing things at the workshop on Biological and Bio-inspired Information Theory.

The first big thing was Nafthali Tishby's program to develop a principled approach to biology and intelligence using a combination of

  • partially observed Markov decision processes,
  • Bayesian networks,
  • rate-distortion theory (a branch of information theory)
  • Bellman's equation (from control theory).

For some details, read my blog article. Very briefly, the idea is that

organisms store and process information about their past to make decisions to achieve goals in the future, and optimizing this process must take into account the price for information storage and processing.

This idea is sort of obvious... and oversimplified: reality is more complex. But the important part is that Tishby and his colleagues have the mathematical tools to study this idea quantitatively by writing software and proving theorems! It's not just chat, it's math.

I think there's a huge amount left to be done here - and that's fine: this is the kind of ambitious synthesis I want to be involved in! If "green mathematics" ever succeeds, it will have to include ideas like this (and much more).

So, I'm going to steer my research in this direction. It's not hard, because I'm already working on control theory, Markov processes and information theory, trying to fit them together in a unified whole.

That's the first big thing.

Comments

  • 1.
    edited November 2014

    The second big thing is that Susanne Still and a grad student of hers have written software that uses similar information-theoretic ideas to make predictions of time series given past data and to estimate the optimal model given constraints on how much information the model can use. Even better, my student Blake Pollard is working with her to apply this software to El Niño data!

    He will start with a simple demonstration just to help her write a paper on this subject. But we may expand this to a larger project: to study Niño and other climate phenomena using ideas from information theory!

    This is very nice because again it's compatible with a lot of things I already want to do, and things the Azimuth Code Project is starting to do... and it begins to connect some of these ideas.

    Here is some reading material suggested by Susanne:

    If you think you're getting bombarded with too many references, don't worry! I will read a bunch of this stuff and explain it on the blog. You just need to read the blog articles. (I know, that's already hard enough!)

    Anyway, I'm very excited.

    Comment Source:The _second_ big thing is that [Susanne Still](http://www2.hawaii.edu/~sstill/) and a grad student of hers have written software that uses similar information-theoretic ideas to make predictions of time series given past data and to estimate _the optimal model given constraints on how much information the model can use_. Even better, my student [[Blake Pollard]] is working with her to apply this software to El Niño data! He will start with a simple demonstration just to help her write a paper on this subject. But we may expand this to a larger project: to study Niño and other climate phenomena using ideas from information theory! This is very nice because again it's compatible with a lot of things I already want to do, and things the Azimuth Code Project is starting to do... and it begins to _connect_ some of these ideas. Here is some reading material suggested by Susanne: * Naftali Tishby, Fernando C. Pereira and William Bialek, [The information bottleneck method](http://www.cnbc.cmu.edu/cns/papers/Tishby-NC-1999.pdf). * Susanne Still, James P. Crutchfield and Christopher J. Ellison, [Optimal causal inference: estimating stored information and approximating causal architecture](http://arxiv.org/abs/0708.1580). * Susanne Still and William Bialek, [How many clusters? an information-theoretic perspective](http://www2.hawaii.edu/~sstill/HowManyClusters.pdf). * Susanne Still, [Information theoretic approach to interactive learning](http://www2.hawaii.edu/~sstill/Still_IL2009_EPL.pdf). * Susanne Still, [Information bottleneck approach to predictive inference](http://www.mdpi.com/1099-4300/16/2/968). (Part of a [special issue of _Entropy_ on the information bottleneck method](http://www.mdpi.com/journal/entropy/special_issues/bottleneck-method).) * Susanne Still, David A. Sivak, Anthony J. Bell and Gavin E. Crooks [The thermodynamics of prediction](http://arxiv.org/abs/1203.3271). If you think you're getting bombarded with too many references, don't worry! I will read a bunch of this stuff and explain it on the blog. You just need to read the blog articles. (I know, that's already hard enough!) Anyway, I'm very excited.
  • 2.

    Hi John

    Is the software Blake Pollard is using the "Sir Isaac" system that Nemenman talks about in Automated adaptive inference of coarse-grained dynamical models in systems biology and associated presentation?

    Or does Susanne Still have a different system?

    Is either system publicly available? I have not been able to find any further information on either Sir Isaac or any system developed by Susanne Still.

    cheers Daniel

    Comment Source:Hi John Is the software Blake Pollard is using the "Sir Isaac" system that Nemenman talks about in [Automated adaptive inference of coarse-grained dynamical models in systems biology](http://arxiv.org/abs/1404.6283) and associated [presentation](http://www.nemenmanlab.org/~ilya/images/9/92/DynamicDays_2014.pdf)? Or does Susanne Still have a different system? Is either system publicly available? I have not been able to find any further information on either Sir Isaac or any system developed by Susanne Still. cheers Daniel
  • 3.
    edited November 2014

    I found Sir Isaac. Is Susanne Still using it or something else?

    Comment Source:I found [Sir Isaac](https://github.com/EmoryUniversityTheoreticalBiophysics/SirIsaac). Is Susanne Still using it or something else?
  • 4.
    edited November 2014

    I'll ask Blake Pollard what system Susanne Still is using. I don't know its name. I think it's something new. She's written a theoretical paper on it, but she wants to include an example of how it works.

    Comment Source:I'll ask Blake Pollard what system Susanne Still is using. I don't know its name. I think it's something new. She's written a theoretical paper on it, but she wants to include an example of how it works.
  • 5.

    She doesn't use any particular software. The code that implements her method is written in C++. I got in touch with her to get the code. She calls the method 'optimal causal inference'. It is uses a recursive information bottleneck algorithm. The information bottleneck approach was introduced by Naftali Tishby.

    Comment Source:She doesn't use any particular software. The code that implements her method is written in C++. I got in touch with her to get the code. She calls the method 'optimal causal inference'. It is uses a recursive information bottleneck algorithm. The information bottleneck approach was introduced by Naftali Tishby.
  • 6.

    Thanks John

    Comment Source:Thanks John
  • 7.

    ... and thanks Blake.

    Comment Source:... and thanks Blake.
Sign In or Register to comment.