I read somewhere about 'the kernel trick' (a machine learning concept used in vector classification), and decided to implement a few major distances (euclidean,

I read somewhere about 'the kernel trick' (a machine learning concept used in vector classification), and decided to implement a few major distances (euclidean, absolute and dot product) and see what they would actually look like. Here are 100 cases of the Iris dataset in a 'euclidean kernel': https://twitter.com/adrnsly/status/455305310488363008/photo/1 https://twitter.com/adrnsly/status/455305310488363008/photo/... Turns out 'the kernel trick' and KNN classifiers can be chained endlessly in a recursive structure (as long as you build and propagate your neighbourhoods right!), which means they can be used as an activation system in a neural network (as opposed to the sigmoid or hypertan functions).