Calculates the softMin of a vector.
Let D be a vector. Then the softMin of D is defined as:
s = exp(-D/sigma^2) / sum( exp(-D/sigma^2) )
The softMin is a way of taking a dissimilarity (distance) vector D and
converting it to a similarity vector s, such that sum(s)==1. If D is an
NxK array, is is treated as N K-dimensional vectors, and the return is
likewise an NxK array. This is useful if D is a distance matrix,
generated by the likes of pdist2.
Note that as sigma->0, softMin's behavior tends toward that of the
standard min function. That is the softMin of a vector D has all zeros
with a single 1 in the location of the smallest value of D. For example,
"softMin([.2 .4 .1 .3],eps)" returns "[0 0 1 0]". As sigma->inf, then
softMin(D,sigma) tends toward "ones(1,n)/n", where n==length(D).
If D contains the squared euclidean distance between a point y and k
points xi, then there is a probabilistic interpretation for softMin. If
we think of the k points representing equal variant gaussians each with
mean xi and std sigma, then the softMin returns the relative probability
of y being generated by each gaussian.
USAGE
M = softMin( D, sigma )
INPUTS
D - NxK dissimilarity matrix
sigma - controls 'softness' of softMin
OUTPUTS
M - the softMin (indexes into D)
EXAMPLE - 1
C = [0 0; 1 0; 0 1; 1 1]; x=[.7,.3; .1 .2];
D = pdist2( x, C ), M = softMin( D, .25 )
EXAMPLE - 2
fplot( 'softMin( [0.5 0.2 .4], x )', [0 5] );
xlabel('sigma'); ylabel('assignments')
See also PDIST2, SOFTMAX
Piotr's Computer Vision Matlab Toolbox Version 2.0
Copyright 2014 Piotr Dollar. [pdollar-at-gmail.com]
Licensed under the Simplified BSD License [see external/bsd.txt]