MATLAB File Help: dml.searchlight

  searchlight searchlight analysis.
 
    DESCRIPTION
    The searchlight uses a sphere with a certain radius and stepsize in order 
    to scan a volume and classify based on elements inside the sphere. 
    Subsequently one may either:
    * Use m.model to return the performance estimates throughout the volume
    * Select the best n spheres and retrieve this subset for the input data. 
      The number nspheres is computed using a nested cross-validation with 
      the specified validator.
 
    EXAMPLE:
 
    X = rand(10,1000); Y = [1 1 1 1 1 2 2 2 2 2]';
 
    Perform searchlight and return classification accuracy computed with a 
    default crossvalidator and mapped back to the standard volume. 
 
    indims = [10 10 10];
    m = dml.searchlight('step',2,'radius',2,'indims',indims,'verbose',true,'stats',{'accuracy'});
    m = m.train(X,Y); r = m.model;
 
    Alternatively, a multidimensional logical mask can be used to specify
    the input dimensions and restrict to a subset of the original volume; the
    original volume will still be used to estimate the searchlight
    spheres:
 
    mymask = true(10,10,10); mymask(:,:,1:5) = false;
    m = dml.searchlight('step',2,'radius',2,'mask',mymask,'verbose',true,'stats',{'accuracy'});
    m = m.train(X(:,find(mymask(:))),Y); r = m.model;
 
    We may also specify a different, irregular, neighbourhood structure
    in conjunction with the input dimensions or a particular mask:
 
    mymask = true(10,10,10); mymask(:,:,1:5) = false;
    nb = sparse(1000,1000); prm = randperm(1e6); nb(prm(1:1000)) = 1; nb = (nb + nb') ~= 0;
    m = dml.searchlight('step',2,'radius',2,'neighbours',nb,'mask',mymask,'verbose',true,'stats',{'accuracy'});
    m = m.train(X(:,find(mymask(:))),Y); r = m.model;
 
    The output of this searchlight analysis can also be used for feature selection.
    In that case, one needs to specify nspheres, containing the numbers of
    spheres to test. The optimal subset of spheres will be determined using
    the specified crossvalidator:
 
    mymask = true(10,10,10);
    m = dml.searchlight('nspheres',[1 2 3 4 5],'step',2,'radius',2,'mask',mymask,'verbose',true,'stats',{'accuracy'});
    m = m.train(X,Y); r = m.model;
 
    Instead of using a cross-validator, one may also use a permutation
    object to determine sphere performance:
 
    indims = [10 10 10];
    p = dml.permutation('stat','accuracy','validator',dml.crossvalidator('mva',dml.svm),'nperm',10,'verbose',true);
    m = dml.searchlight('step',3,'radius',2,'indims',indims,'verbose',true,'validator',p);
    m = m.train(X,Y); r = m.model;
 
    DEVELOPER
    Marcel van Gerven (m.vangerven@donders.ru.nl)

Superclasses	dml.method
Sealed	false
Construct on load	false

center	only selects the feature in the centre of the sphere if true
centers	centers of each sphere
compact	save validators if compact = false
exclude	if false then only the center within a sphere is required to be part of the optional mask
indims	dimensions of the input data (excluding the trial dim and time dim in time series data)
mask	optional logical mask of size indims (input features are only those in the mask)
neighbours	a sparse adjacency matrix specifying the neighbourhood structure for irregular data (don't use in conjunction with mask)
nspheres	select the union of the best nspheres spheres;
original	in original space
performance	performance as function of number of selected spheres (or sphere centers)
radius	radius of the hypersphere in terms of array elements (diameter will be 1 + 2 * radius).
restart	when false, starts at the previously learned parameters; needed for online learning and grid search
spheres	the features belonging to each sphere
stats	the statistics to save in value (e.g. {'accuracy','binomial'}); empty stats field will just call validator.statistic
step	stepsize in terms of array elements
subset	the used feature subset
validator	the validator used to determine the final used feature subset
value	evaluation metric per sphere
verbose	whether or not to generate diagnostic output

	estimate_spheres
	model	return performance mapped back to original space
	test
	train