Difference between revisions of "Randomness, Structure and Causality - Agenda"
From Santa Fe Institute Events Wiki
Line 209: | Line 209: | ||
Still, Suzanne (sstill@hawaii.edu)<br> | Still, Suzanne (sstill@hawaii.edu)<br> | ||
University of Hawaii at Manoa | University of Hawaii at Manoa | ||
− | <br><br> | + | <br> |
− | Links: | + | <br> |
+ | The principles of statistical mechanics and information theory play an important role in learning and have inspired both theory and the design of numerous machine learning algorithms. The new aspect in this paper is a focus on integrating feedback from the learner. A quantitative approach to interactive learning and adaptive behavior is proposed, integrating model- and decision-making into one theoretical framework. This paper follows simple principles by requiring that the observer’s world model and action policy should result in maximal predictive power at minimal complexity. Classes of optimal action policies and of optimal models are derived from an objective function that reflects this trade-off between prediction and complexity. The resulting optimal models then summarize, at different levels of abstraction, the process’s causal organization in the presence of the learner’s actions. A fundamental consequence of the proposed principle is that the learner’s optimal action policies balance exploration and control as an emerging property. Interestingly, the explorative component is present in the absence of policy randomness, i.e. in the optimal deterministic behavior. This is a direct result of requiring maximal predictive power in the presence of feedback. | ||
+ | <br> | ||
+ | <br> | ||
+ | Links: [[http://arxiv.org/abs/0709.1948]] | ||
---- | ---- | ||
'''Measuring the Complexity of Psychological States''' | '''Measuring the Complexity of Psychological States''' |
Revision as of 18:34, 17 December 2010
Workshop Navigation |
Abstracts
Effective Complexity of Stationary Process Realizations
Ay, Nihat (nay@mis.mpg.de)
SFI & Max Planck Institute
Links: [[1]]
Learning Out of Equilibrium
Bell, Tony (tony@salk.edu)
UC Berkeley
Inspired by new results in non-equilibrium statistical mechanics, we define a new kind of state-machine that can be used to model time series. The machine is deterministically coupled to the inputs unlike stochastic generative models like the Kalman filter and HMM’s. The likelihood in this case is shown to be a sum of local time likelihoods. We introduce a new concept, second-order-in-time stochastic gradient, which derives from the time derivative of the likelihood, showing that the latter decomposes into a ‘work’ term, a ‘heat’ term and a term describing time asymmetry in the state machine’s dynamics. This motivates the introduction of a new time-symmetric likelihood function for time series. Our central result is that the time derivative of this is an average sum of forward and backward time ‘work’ terms, in which all partition functions, which plague Dynamic Bayesian Networks, have cancelled out. We can now do tractable time series density estimation with arbitrary models, without sampling. This is a direct result of doing second-order-in-time learning with time-symmetric likelihoods. A model is proposed, based on parameterised energy-based Markovian kinetics, with the goal of learning (bio)chemical networks from data, and taking a step towards understanding molecular-level energy-based self-organisation.
Links:
The Transmission of Sense Information
Bergstrom, Carl (cbergst@u.washington.edu)
SFI & University of Washington
Biologists rely heavily on the language of information, coding, and transmission that is commonplace in the field of information theory developed by Claude Shannon, but there is open debate about whether such language is anything more than facile metaphor. Philosophers of biology have argued that when biologists talk about information in genes and in evolution, they are not talking about the sort of information that Shannon’s theory addresses. First, philosophers have suggested that Shannon’s theory is only useful for developing a shallow notion of correlation, the so-called ‘‘causal sense’’ of information. Second, they typically argue that in genetics and evolutionary biology, infor- mation language is used in a ‘‘semantic sense,’’ whereas semantics are deliber- ately omitted from Shannon’s theory. Neither critique is well-founded. Here we propose an alternative to the causal and semantic senses of information: a transmission sense of information, in which an object X conveys information if the function of X is to reduce, by virtue of its sequence properties, uncertainty on the part of an agent who observes X. The transmission sense not only captures much of what biologists intend when they talk about information in genes, but also brings Shannon’s theory back to the fore. By taking the view- point of a communications engineer and focusing on the decision problem of how information is to be packaged for transport, this approach resolves several problems that have plagued the information concept in biology, and highlights a number of important features of the way that information is encoded, stored, and transmitted as genetic sequence.
Links: [[2]]
Optimizing Information Flow in Small Genetic Networks
Bialek, William (wbialek@Princeton.EDU)
Princeton University
Links: [[3]]
To a Mathematical Theory of Evolution and Biological Creativity
Chaitin, Gregory (gjchaitin@gmail.com)
IBM Watson Research Center
We present an information-theoretic analysis of Darwin’s theory of evolution, modeled as a hill-climbing algorithm on a fitness landscape. Our space of possible organisms consists of computer programs, which are subjected to random mutations. We study the random walk of increasing fitness made by a single mutating organism. In two different models we are able to show that evolution will occur and to characterize the rate of evolutionary progress, i.e., the rate of biological creativity.
Links: File:Darwin.pdf
Framing Complexity
Crutchfield, James (chaos@cse.ucdavis.edu)
SFI & UC Davis
Is there a theory of complex systems? And who should care, anyway?
Links: [[4]]
The Vocabulary of Grammar-Based Codes and the Logical Consistency of Texts
Debowski, Lukasz (ldebowsk@ipipan.waw.pl)
Polish Academy of Sciences
We will present a new explanation for the distribution of words in
natural language which is grounded in information theory and inspired
by recent research in excess entropy. Namely, we will demonstrate a
theorem with the following informal statement: If a text of length
describes independent facts in a repetitive way then the
text contains at least different words. In the
formal statement, two modeling postulates are adopted. Firstly, the
words are understood as nonterminal symbols of the shortest
grammar-based encoding of the text. Secondly, the text is assumed to
be emitted by a finite-energy strongly nonergodic source whereas the
facts are binary IID variables predictable in a shift-invariant
way. Besides the theorem, we will exhibit a few stochastic processes
to which this and similar statements can be related.
Links: [[5]] and [[6]]
Prediction, Retrodiction, and the Amount of Information Stored in the Present
Ellison, Christopher (cellison@cse.ucdavis.edu)
Complexity Sciences Center, UC Davis
We introduce an ambidextrous view of stochastic dynamical systems, comparing their forward-time and reverse-time representations and then integrating them into a single time-symmetric representation. The perspective is useful theoretically, computationally, and conceptually. Mathematically, we prove that the excess entropy--a familiar measure of organization in complex systems--is the mutual information not only between the past and future, but also between the predictive and retrodictive causal states. Practically, we exploit the connection between prediction and retrodiction to directly calculate the excess entropy. Conceptually, these lead one to discover new system invariants for stochastic dynamical systems: crypticity (information accessibility) and causal irreversibility. Ultimately, we introduce a time-symmetric representation that unifies all these quantities, compressing the two directional representations into one. The resulting compression offers a new conception of the amount of information stored in the present.
Links: [[7]]
Complexity Measures and Frustration
Feldman, David (dave@hornacek.coa.edu)
College of the Atlantic
In this talk I will present some new results applying complexity
measures to frustrated systems, and I will also comment on some
frustrations I have about past and current work in complexity
measures. I will conclude with a number of open questions and ideas
for future research.
I will begin with a quick review of the excess entropy/predictive information and argue that it is a well understood and broadly applicable measure of complexity that allows for a comparison of information processing abilities among very different systems. The vehicle for this comparison is the complexity-entropy diagram, a scatter-plot of the entropy and excess entropy as model parameters are varied. This allows for a direct comparison in terms of the configurations' intrinsic information processing properties. To illustrate this point, I will show complexity-entropy diagrams for: 1D and 2D Ising models, 1D Cellular Automata, the logistic map, an ensemble of Markov chains, and an ensemble of epsilon-machines.
I will then present some new work in which a local form of the 2D excess entropy is calculated for a frustrated spin system. This allows one to see how information and memory are shared unevenly across the lattice as the system enters a glassy state. These results show that localised information theoretic complexity measures can be usefully applied to heterogeneous lattice systems. I will argue that local complexity measures for higher-dimensional and heterogeneous systems is a particularly fruitful area for future research.
Finally, I will conclude by remarking upon some of the areas of
complexity-measure research that have been sources of frustration.
These include the persistent notions of a universal "complexity at
the edge of chaos," and the relative lack of applications of
complexity measures to empirical data and/or multidimensional systems.
These remarks are designed to provoke dialog and discussion about
interesting and fun areas for future research.
Links: File:Afm.tri.5.pdf and File:CHAOEH184043106 1.pdf
Complexity, Parallel Computation and Statistical Physics
Machta, Jon (machta@physics.umass.edu)
SFI & University of Massachusetts
Links: [[8]]
Crypticity and Information Accessibility
Mahoney, John (jmahoney3@ucmerced.edu)
UC Merced
We give a systematic expansion of the crypticity--a recently introduced measure of the inaccessibility of a stationary process's internal state information. This leads to a hierarchy of k-cryptic processes and allows us to identify finite-state processes that have infinite crypticity--the internal state information is present across arbitrarily long, observed sequences. The crypticity expansion is exact in both the finite- and infinite-order cases. It turns out that k-crypticity is complementary to the Markovian finite-order property that describes state information in processes. One application of these results is an efficient expansion of the excess entropy--the mutual information between a process's infinite past and infinite future--that is finite and exact for finite-order cryptic processes.
Links: [[9]]
Automatic Identification of Information-Processing Structures in Cellular Automata
Mitchell, Melanie (mm@cs.pdx.edu)
SFI & Portland State University
Cellular automata have been widely used as idealized models of natural spatially-extended dynamical systems. An open question is how to best understand such systems in terms of their information-processing capabilities. In this talk we address this question by describing several approaches to automatically identifying the structures underlying information processing in cellular automata. In particular, we review the computational mechanics methods of Crutchfield et al., the local sensitivity and local statistical complexity filters proposed by Shalizi et al., and the information theoretic filters proposed by Lizier et al. We illustrate these methods by applying them to several one- and two-dimensional cellular automata that have been designed to perform the so-called density (or majority) classification task.
Phase Transitions and Computational Complexity
Moore, Cris (moore@cs.unm.edu)
SFI & University of New Mexico
Links:
Statistical Mechanics of Interactive Learning
Still, Suzanne (sstill@hawaii.edu)
University of Hawaii at Manoa
The principles of statistical mechanics and information theory play an important role in learning and have inspired both theory and the design of numerous machine learning algorithms. The new aspect in this paper is a focus on integrating feedback from the learner. A quantitative approach to interactive learning and adaptive behavior is proposed, integrating model- and decision-making into one theoretical framework. This paper follows simple principles by requiring that the observer’s world model and action policy should result in maximal predictive power at minimal complexity. Classes of optimal action policies and of optimal models are derived from an objective function that reflects this trade-off between prediction and complexity. The resulting optimal models then summarize, at different levels of abstraction, the process’s causal organization in the presence of the learner’s actions. A fundamental consequence of the proposed principle is that the learner’s optimal action policies balance exploration and control as an emerging property. Interestingly, the explorative component is present in the absence of policy randomness, i.e. in the optimal deterministic behavior. This is a direct result of requiring maximal predictive power in the presence of feedback.
Links: [[10]]
Measuring the Complexity of Psychological States
Tononi, Guilio (gtononi@wisc.edu)
University of Michigan
Links:
Ergodic Parameters and Dynamical Complexity
Vilela-Mendes, Rui (vilela@cii.fc.ul.pt)
University of Lisbon
Using a cocycle formulation, old and new ergodic parameters beyond the
Lyapunov exponent are rigorously characterized. Dynamical Renyi entropies
and fluctuations of the local expansion rate are related by a generalization
of the Pesin formula.
How the ergodic parameters may be used to characterize the complexity of
dynamical systems is illustrated by some examples: Clustering and
synchronization, self-organized criticality and the topological structure of
networks.
Links: [[11]]
Hidden Quantum Markov Models and Non-adaptive Read-out of Many-body States
Wiesner, Karoline (k.wiesner@bristol.ac.uk)
University of Bristol
Links: [[12]]