Code

Please see the AMA code release at the BurgeLab GitHub repository for our most update-to-date code:

                                                          http://github.com/burgelab


Accuracy Maximization Analysis (AMA) - Matlab and C code implementation

Relevant publications: Please cite the following papers if the code on this page

                                    or in the GitHub repository contributes to your research


               • Burge J, Jaini P (2017). Accuracy Maximization Analysis for sensory-perceptual tasks: Computational

                  improvements, filter robustness, and coding advantages for scaled additive noise.  PLoS Computational

                  Biology, 13(2): e1005281. doi:10.1371/journal.pcbi.1005281 [ pdf ]

               • Jaini P, Burge J (2017). Linking normative models of natural tasks with descriptive models of neural

                  response.  bioarXiv, doi:https://doi.org/10.1101/158741 [ pdf ]

               • Burge J, Geisler WS (2015). Optimal speed estimation in natural image movies predicts human

                  performance. Nature Communications, 6: 7900, 1-11 doi:10.1038/ncomms8900 [ pdf ]

               • Burge J, Geisler WS (2011). Optimal defocus estimation in individual natural images. Proceedings of

                  the National Academy of Sciences, 108 (40): 16849-16854 [ pdf ]

               • Geisler WS, Najemnik J, Ing AD (2009). Optimal stimulus encoders for natural tasks. Journal of

                   Vision, 9(13):17, 1-16 [ pdf | correction ]


Download: AMAengine.zip ( more info: AMAengine_ReadMe.rtf )


Description: Computes the posterior probability of a category given the filter response(s) to the stimulus


Setup:          (1) Download and unzip AMAengine.zip

                    (2) Put all five unzipped files in the current folder and/or in a folder on the Matlab path

                    (3) At the command prompt type: >> mex AMAengine.cpp                    

                               ...OR, if it errors out, type: >> mex -compatibleArrayDims AMAengine.cpp

                    (4) At the command prompt type: >> help AMAengine

                    (5) At the command prompt type: >> AMAengineTest;


Use: AMA requires computing the posterior probability of each stimulus category given the filter responses

        for every stimulus in the training set


        To use AMA, a minimization routine and an objective function must be chosen

        Note that the minimization routine and objective function are not fundamental to the AMA procedure

        To get you started, an example minimization routine and an example objective function are provided

        However, other minimization routines and/or objective functions may better suit your needs


        The example below uses Matlab’s fmincon.m to find AMA filters that minimize the average relative

        entropy of the posterior probability distributions over the categories, across all stimuli in the training set.


        NOTE! Please download the code from github link above for the most recent and best functioning code. The code pasted

        below is best thought of as very detailed pseudo code, rather than code that should be copied and used as is.


        Troubleshooting (Matlab R2011a, R2011b, R2012a and Xcode 4.2, Xcode 4.3)

        Any combination of these versions of Mathworks & Apple software may prevent Matlab from finding the C++ header files

        To fix the issue, go to http://www.mathworks.com/support/solutions/en/data/1-FR6LXJ/ and follow the instructions

(1) Select stimulus:                                               s(k,l)

(2) Noisy response due to that s(k,l):                      N(r(k,l),sigma(k,l))

(3) Noisy responses due to other stims in same Ctg: N(r(k,j),sigma(k,j))

(4) Noisy responses due to other stims in other Ctgs: N(r(i,j),sigma(i,j))

(5) Likelihood of response r(k,l) from s(k,l)

(6) Likelihood of response r(k,l) from s(k,j) and s(i,j)

(7) Posterior probability of correct category X(k) given s(k,l)

(8) Repeat for all other stimuli in the training set

Logic:

All code is published under the GNU General Public License: this software is free; you can

redistribute it and/or modify it under the terms of the GNU General Public License (GPLv3)

Please report bugs to jburge@mail.cps.utexas.edu with subject jburge.cps.utexas.edu/Code

%%% OBJECTIVE FUNCTION %%%

function C = objFunc(f,s,ctgInd,X,cstType,rMax,fano,v0)

% function C = objFunc(f,s,ctgInd,X,cstType,rMax,fano,v0)

%

% objective function for AMA

%

% f:       filters. vector magnitude of each  filter  must equal 1         [ nDim x nF   ]

% s:       stimuli. vector magnitude of each stimulus must equal 1         [ nStm x nDim ]

% ctgInd:  category indices of each stimulus (must be integer valued)      [ nStm x 1    ]

%          NOTE! number of unique values in ctgInd number of levels in X  

% X:       latent variable values                                          [    1 x nLvl ]

% cstType: cost type                                                       [    1 x nLvl ]

% rMax:       max mean response (i.e. if cos similarity btwn f and s = 1.0    [   scalar    ]

% fano:       fano factor... how neural noise variance scales w mean response [   scalar    ]

% v0

% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% C:       average cost to minimize                                        [    1 x 1    ]


% INPUT HANDLING

if min(ctgInd) == 0, error(['objFunc: WARNING! min(ctgInd) must be equal 1']); end; % error check


% MEAN FILTER RESPONSE(S) TO ALL STIMULI

r          =  stim2resp(f,s,rMax);           

% STDDEV OF FILTER RESPONSE(S) TO ALL STIMULI

sigma      =  resp2sigma(r,fano,v0);      

% POSTERIOR PROB AT CORRECT CTG (pp [nStm x 1]) & ALL CTGs (ppAll [nStm x nLvl]) FOR ALL STIMULI    

[pp,ppAll] =  AMAengine(r,r,sigma,ctgInd);  


% NUMBER OF STIMS

nStm = size(s,1);


% NUMBER OF STIMS

if strcmp(cstType,'KLD')

  % AVERAGE COST ACROSS STIMULI (KL DIVERGENCE BTWN IDEALIZED & ACTUAL POSTERIOR PROBABILITY DSTBs)

    Call   = -log2(pp);           % average relative entropy in bits

elseif strcmp(cstType,'L0N')

  % AVERAGE COST ACROSS STIMULI

    Call   = 1 - pp;              % posterior map is optimal for L0 norm cost function

elseif strcmp(cstType,'L2N')

  % POSTERIOR MEAN (optimal estimate for L2norm cost function)

    Xhat   = sum(bsxfun(@times,ppAll,X),2);       

  % COST FOR EACH STIMULUS

    Call   = ((Xhat - X(ctgInd)').^2);    

end

% AVERAGE COST ACROSS STIMULI

C      = sum(Call,1)./nStm;

%%% MINIMIZATION ROUTINE %%%

>> [f] = fmincon(@(f) objFunc(f,s,ctgInd,X,cstType,rMax,fano,v0),f0,[],[],[],[],[],[],@(f) fConEq(f));

%%% CONSTRAINT FUNCTION %%%

function [c,ceq] = fConEq(f)

c = [];

ceq = sum(f.^2,1)-1;                         % vector magnitude of filters must equal 1.0

%%%  RESPONSE FUNCTION  %%%

function [r] = stim2resp(f,s,rMax)

r = rMax.*(s*f);                             % mean filter response        [ nStm x nF ]

%%%   SIGMA   FUNCTION  %%%

function [sigma] = resp2sigma(r,fano,v0)

sigma = sqrt( fano.*abs(r) + v0 );           % stddev of filter response   [ nStm x nF ]