Merge pull request translationalneuromodeling#155 from translationaln…

…euromodeling/development

Release v5.1.2

CHANGELOG.md

@@ -1,8 +1,12 @@
 # Changelog
 TAPAS toolbox 
 
-## [5.1.1] 2021-06-22
+## [5.1.2] 2021-09-06
+
+### Fixed
+- TOOLS: Fixed a bug in the computation of WAIC (function [tapas_waic.m](tools/tapas_waic.m)). Added new function to compute log-exp_sum (function [tapas_logsumexp.m](tools/tapas_logsumexp.m)).
 
+## [5.1.1] 2021-06-22
 
 ### Fixed
 - Added references to TAPAS paper (in collection and toolboxes)

README.md

@@ -1,6 +1,6 @@
 ![TAPAS Logo](misc/TapasLogo.png?raw=true "TAPAS Logo")
 
-*Version 5.1.1*
+*Version 5.1.2*
 
 T  A  P  A  S - Translational Algorithms for Psychiatry-Advancing Science.
 ========================================================================
@@ -15,9 +15,10 @@ DESCRIPTION
 -----------
 
 TAPAS is a collection of algorithms and software tools developed by the
-Translational Neuromodeling Unit (TNU, Zurich) and collaborators. The goal of
-these tools is to support clinical neuromodeling, particularly computational
-psychiatry, computational neurology, and computational psychosomatics.
+Translational Neuromodeling Unit (TNU, Zurich) and collaborators. These 
+tools are intended to support the development of computational assays 
+(Translational Neuromodeling) and their clinical application in 
+Computational Psychiatry, Neurology and Psychosomatics.
 
 Currently, TAPAS includes the following packages:
 

misc/log_tapas.txt

@@ -1,3 +1,4 @@
+5.1.2   https://www.tapas.tnu-zurich.com/examples_v5.0.0.zip   24c1248d3054f025b853b5ab7ce08a1a
 5.1.1   https://www.tapas.tnu-zurich.com/examples_v5.0.0.zip   24c1248d3054f025b853b5ab7ce08a1a
 5.1.0   https://www.tapas.tnu-zurich.com/examples_v5.0.0.zip   24c1248d3054f025b853b5ab7ce08a1a
 5.0.0   https://www.tapas.tnu-zurich.com/examples_v5.0.0.zip   24c1248d3054f025b853b5ab7ce08a1a

tools/tapas_logsumexp.m

@@ -0,0 +1,42 @@
+function [y_sum,y_mean] = tapas_logsumexp(x)
+
+
+%% -------------------------------------------------------------------------------------------
+% [y_sum,y_mean] = tapas_logsumexp(x) takes the values in x, exponates
+% them, then takes the sum over the column, and finally applies the natural logarithm.
+% The calculation uses the "log-sum-exp" trick: See e.g. http://gregorygundersen.com/blog/2020/02/09/log-sum-exp/
+% The function also returns the log-mean-exp.
+%---------------------------------------------------------------------------------------------
+% INPUT:
+%       x       - A column vector or matrix of values. All computations are
+%       made along the direction of a column.
+% 
+% Optional:
+%
+%--------------------------------------------------------------------------------------------
+% OUTPUT: 
+%       y_sum   - The log-sum-exp of all columns of x.
+%       y_mean  - The log-mean-exp of all columns of x.
+%           
+% Author:  Jakob Heinzle, TNU, UZH & ETHZ - April, 2021
+%
+% REVISION LOG:
+%
+%      Jakob Heinzle, 2021/04/16: new function 
+%
+%%
+
+sz = size(x);
+
+if numel(sz)~=2
+    error('Input x needs to be a matrix of 2 dimensions');
+end
+
+max_x = max(x); %compute maximum of each column
+y_sum = max_x + log(sum(exp(x-ones(sz(1),1)*max_x)));
+
+if nargout==2
+y_mean = y_sum-log(sz(1)); % compute mean if necessary.
+end
+
+return;

tools/tapas_waic.m

@@ -1,4 +1,4 @@
-function [waic, accuracy] = tapas_waic(llh)
+function [waic, lppd] = tapas_waic(llh)
 %% Computes the Watanabe-Akaike Information criterion
 %
 % Input
@@ -7,10 +7,10 @@
 %                       the number of samples.
 %
 % Output
-%       waic        -- The Watanaba AIC usin the methods in [1].
-%       accuracy    -- The expected log likelihood of the model.
+%       waic        -- The Watanabe AIC usin the methods in [1].
+%       lppd    -- The log pointwise predictive density of the model.
 %
-% The WAIC can be computed as the accuracy - penalization, where the
+% The WAIC can be computed as the lppd - penalization, where the
 % penalization is the sum of the variance of the log likelihood, i.e., the
 % gradient of the Free energy.
 %
@@ -21,13 +21,21 @@
 % aponteeduardo@gmail.com
 % copyright (C) 2019
 %
+% REVISION LOG:
+%
+%      Jakob Heinzle, 2021/04/16: changed computation of accuracy to be exactly log
+%      pointwise predictive density as in Gelman et al.
+%
+%%
+
+s = size(llh,2);
 
 % Estimator of the accuracy
-accuracy = sum(mean(llh, 2));
+lppd = sum(tapas_logsumexp(llh')-log(s));
 
 % Estimator of the variance
 penalization = sum(var(llh, [], 2));
 
-waic = accuracy - penalization;
+waic = lppd - penalization;
 
 end