Skip to main content

Table 1 Learning principles

From: Unifying generative and discriminative learning principles

  

prior knowledge

  

non Bayesian

Bayesian

 

generative

ML

MAP

objective

hybrid

GDT

PGDT

 

discriminative

MCL

MSP

  1. The table shows six established learning principles that can be grouped by their objective as being generative, hybrid, or discriminative and utilization of prior knowledge with the two possibilities non Bayesian and Bayesian. The four elementary learning principles are the generative, non Bayesian maximum likelihood (ML) learning principle, the generative, Bayesian maximum a posteriori (MAP) learning principle, the discriminative, non Bayesian maximum conditional likelihood (MCL) learning principle, and the discriminative, Bayesian maximum supervised posterior (MSP) learning principle. The hybrid learning principles which interpolate between generative and discriminative learning principles are the non Bayesian generative-discriminative trade-off (GDT) learning principle and the penalized generative-discriminative trade-off (PGDT) learning principle.