Softmax normalization
maps real scores to a distribution of positive values summing to 1.
symbols, variables and units
z_i: dimensionless logit; p_i: probability form output.
applicable conditions and boundaries
numerical calculations should subtract the maximum logit to avoid overflow; the output may not be calibrated.
formula source code
The following is a copyable LaTeX expression.
p_i=\frac{e^{z_i}}{\sum_j e^{z_j}}Reference and Extended Learning
MIT OpenCourseWare · Introduction to Algorithms ↗is organized according to model definition and assumptions. Please check actual conditions and original literature before engineering, research and clinical use.
softmaxMachine Learning