Melvin's digital garden

Machine learning for Big Data analysis

speaker: Kheng Cheng Wai, UTAR event: Big Data Workshop, UTAR Kampar, Malaysia ** deviations are noise from empirical data ** empirical risk minimization does not differentiate between a simple model and a complex model ** VC dimension the number of ways the model can partition the data VC theory can be used to avoid over fitting ** SVM maximizes the margin only a small number of support vectors needed to find the optimal solution relax the contrained optimization problem, solve using quadratic programming ** Gaussian process distribution of models, find the expected value each data point is the expected value of a gaussian distribution training process is to select the hyperparameters to maximize the likelihood O(n^3) need to choose the kernel function as well, similar to SVM prediction is O(n^2 lg n) slower than SVM

Links to this note