4000 handwritten digits, 10 classes (one, two, ..., nine, zero). Each digit is a 16x16 gray scale image. The number of digits per class is intentionally unbalanced: class: one, two, ..., nine, zero #images: 213 129 100 754 970 275 585 166 353 455 The dataset in gzipped plain text file is in skewC.zip and skewX.zip. Distance: The distance between two digits are pixel-wise L2 distance, i.e. d(x,y) = \sqrt( \sum_{i=1}^256 (x_i - y_i)^2 ) . Graph: each digit is a node. The graph is fully connected. The weight between two nodes is w(x,y) = exp(- ( d(x,y)/380 )^2 ), i.e. all 256 dimensions share the same variance 380^2.