machine learning - Are Gaussian clusters linearly separable? -
imagine have 2 gaussian probability distributions in two-dimensions first centered @ (0,1) , second @ (0,-1). (for simplicity, assume have same variance.) can 1 consider clusters of data points sampled these 2 gaussians linearly separable?
intuitively, it's clear boundary separating 2 distributions linear, namely abscissa in our case. however, formal requirement linear separability convex hulls of clusters not overlap. cannot case gaussian-generated clusters since underlying probability distributions pervade of r^2 (albeit negligible probabilities far away mean).
so, gaussian-generated clusters linearly separable? how can 1 reconcile requirement of convex hulls fact straight line conceivable "boundary"? or, perhaps, boundary ceases linear once non-equal variances come in pictures?
the gaussian cluster instances might separable or not. depends on outcome, not on process generating it.
linear separability can defined a existence of plane separating 2 sets of points, such 1 set of points entirely on 1 side of plane, , other set of points entirely on other side of plane.
take specific gaussian distributions. possible generated 2 linearly-seperable sets (either @ abscissa or not). however, probability 1, if variance non-zero, , let processes generate enough points, result not linearly separable.
so, again, question of outcome, not of process.
Comments
Post a Comment