a=loss(lossType,param)
calculates the difference between the input X and the ouput Y depending
on the specified loss type (for further information see below).
The loss can be calculated in two ways. The first is to it by training,
the second is to call the function with a data object as first
parameter (loss(d,loss_type,param)). The results are stored in the Y
part of the data object.
Attributes (with defaults):
type='class_loss' -- type of loss (class_loss,linear_loss...)
param=[] -- used parameters (can also be empty)
Methods:
train,test,calc
LOSS | PARAMETERS & DESCRIPTION
-------------------------------------------------------------------
class_loss -- zero/one loss, L(x,y)=1 if x=y, 0 otherwise
confusion_matrix -- matrix of [true-pos, false-pos; false-neg, true-neg]
epsilon_loss -- L(x,y)= |x-y|, if |x-y|>epsilon, 0 otherwise
linear_loss -- 1-norm, L(x,y)=|x-y|
one_class_loss -- for one-class, e.g novelty detection, etc.
quadratic_loss -- 2-norm, L(x,y)=|x-y|_2^2
roc -- receiver/operator characteristic
roc50 -- receiver/operator characteristic, first n fps
sensitivity -- tp/(tp+fn)
specificity -- tn/(fp+tn)
kernel -- loss derived from kernel matrix (param)
-- inner products in 'loss' space between examples.
alignment -- L(x,y)= sum(sum( (x*x') .* (y*y'))) / normalization