Quantum Field Theory based model-independent learning of continuous distributions; following Nemenman and Bialek, PRE, 2002
Contributors: Ilya Nemenman
(c) Ilya Nemenman, 2000-2014
The algorithm behind this code is described in
Nemenman, I. & Bialek, W. Occam factors and model independent Bayesian learning of continuous distributions. Physical Review E 65, 026137 (2002).
The code itself was written in 2000-2001, with some edits in 2005, to support the analysis in the PRE paper. It was only released in 2014 due to an oversight.
The software is written in Matlab.
Functions/fiiles in package: 0. bcsn.m -- main routine; given the data and auxiliary parameters, performs optimization over the smoothness scale and find the a posteriori optimal probability distribution. 1. qclass.m -- Solving for classical solution of the learning problem (as in BCS'96), given histogramed data and the chosen smoothness scale 2. QNcorrelator.m -- Negative action of Q(x_i), i=1..N. refer to Bialek et.al. 1996; needs the classical solution, the data samples, and the chosen smoothness scale. 3. Rfunctdet.m -- calculates the log of R, functional determinant, for the given classical solution and the smoothness scale. 4. action.m -- calculates the action for minimization over l, the smoothness scale, given the data and the smoothness scale 5. cdiffl.m and cdiffr.m -- cyclic differences functions 6. histQ.m -- histogramming data as needed for the algorithm to work.