Outline • Classification – continued • Parameter estimation Assumptions • Suppose that there are c categories – {1, 2, ....., c} • The prior probability and class conditional density are known • There are a possible actions – {1, 2, ....., a} • Loss function (i | j} describe the loss incurred for taking action i when the state of nature is j 5/29/2016 Visual Perception Modeling 2 Loss Function • The expected loss function given a particular observation x c R( i | x) ( i | j ) P( j | x) j 1 • The overall risk R R( ( x) | x) p( x)dx 5/29/2016 Visual Perception Modeling 3 Bayes Decision Rule • To minimize the overall risk, compute the conditional risk and select the action for which the conditional risk is minimum c R( i | x) ( i | j ) P( j | x) j 1 – The resulting minimum overall risk is called the Bayes risk, which is the best performance 5/29/2016 Visual Perception Modeling 4 Minimum-Error-Rate Classification • Zero-one loss • For minimum error rate, – Decide 1 if P(1 | x) > P(2 | x) – This is the Bayes decision rule 5/29/2016 Visual Perception Modeling 5 Discriminant Functions • The classifier is said to assign a feature vector x to class i if – gi(x) > gj(x) for all j i – This can be viewed as a network – If f(.) is a monotonically increasing function, f(g(x)) and g(x) as discriminant function will give the same classification result 5/29/2016 Visual Perception Modeling 6 Decision Regions • The effect of decision rule is to divide the feature space into c decision regions – R1, R2, ...., Rc – The regions are separated by decision boundaries 5/29/2016 Visual Perception Modeling 7 Normal Density • Gaussian density p ( x) 1 (2 ) d / 2 | Σ |1/ 2 1 t 1 exp ( x μ ) ( x μ ) 2 – Properties • • • • 5/29/2016 Mean Variance Entropy Central limit theorem Visual Perception Modeling 8 Discriminant Functions for Normal Density • Minimum error rate classification for normal density 1 d t 1 g i ( x) ( x i ) i ( x i ) ln( 2 ) 2 2 1 ln(| i |) ln( P( i )) 2 5/29/2016 Visual Perception Modeling 9 Normal Density – Cont. • Case I: – i = 2 I – Linear discriminant function – Minimum distance classifier as a special case • Template matching 5/29/2016 Visual Perception Modeling 10 Normal Density – Cont. • Case II – i = – Mahalanobis distance ( x i ) ( x i ) T 1 – The resulting discriminant function is also linear 5/29/2016 Visual Perception Modeling 11 Normal Density – Cont. • Case III – i arbitrary – Hyper-quadric 5/29/2016 Visual Perception Modeling 12 Bayes Decision Theory – Discrete Features • In this case, a feature can only take one of the m discrete values v1, ....., vm • To minimize the overall risk, select the action that minimizes arg min R(i | x) * i 5/29/2016 Visual Perception Modeling 13 Parameter Estimation • We could design an optimal classifier if we knew the prior probabilities and the classconditional densities – Unfortunately, in pattern recognition applications we rarely have this kind of complete knowledge about the probabilistic structure of the problem • Training data – Some vague, general knowledge about the problem – A number of design samples 5/29/2016 Visual Perception Modeling 14 Parameter Estimation – cont. • Two approaches – Parameter estimation • Estimate the parameters of the unknown probabilities and probability densities – Non-parametric procedures • Multi-layer perceptrons and in general neural networks • Fisher linear discriminant function 5/29/2016 Visual Perception Modeling 15 Parameter Estimation – cont. • Parameter estimation – Maximum-likelihood approach • Parameters as quantities whose values are fixed but unknown • The best estimate of their value is the one that maximizes the probability of obtaining the samples – Bayesian learning • Parameters are random variables with known prior distribution • Observations convert the prior into posteriori 5/29/2016 Visual Perception Modeling 16 Maximum-Likelihood Estimation • The general principle – Log-likelihood • Gaussian cases – Unknown – Unknown and 5/29/2016 Visual Perception Modeling 17 Bayesian Estimation • Class-conditional densities • Parameter Distribution • Gaussian case – Univariate case – Multivariate case 5/29/2016 Visual Perception Modeling 18 Bayesian Estimation – cont. • General theory 5/29/2016 Visual Perception Modeling 19