Figure 1

Illustration of the unified generative-discriminative learning principle. The plots show a projection of the simplex β onto the (β₀, β₁)-plane and the corresponding learning principles for the specific weights encoded by colors. Figure 1(a) shows the general interpretation of the simplex where the points (0, 1), (0, 0.5), (1, 0), and (0.5, 0) refer to the ML, MAP, MCL, and MSP learning principle, respectively, while the lines β₁ = 1 - β₀ and β₁ = 0.5 - β₀ refer to the GDT and PGDT learning principle, respectively. Figure 1(b) shows the interpretation of the unified generative-discriminative learning principle for a conjugate prior that satisfies the condition of equation (11). In this case, each point on the abscissa (β₀-axis) and ordinate (β₁-axis) refers to the MSP and MAP learning principle, respectively, using the prior in a weighted version . The simplex colored in gray corresponds to the MSP learning principle using the weighted posterior as prior for the parameter vector λ.