search

originrose / cortex

Machine learning in Clojure
Eclipse Public License 1.0
1.27k stars 111 forks source link

Tensor losses #255

Closed cnuernber closed 6 years ago

cnuernber commented 6 years ago

Losses are redone using tensors and thus somewhat simplified. This opens the door to removing a good bit of the legacy math subsystem.

cnuernber commented 6 years ago

@poojaramesh : I explored the effect of centerloss in the mnist-example (center-loss-exploration branch I think).

With the change to use the hypersphere, the centroids end up more or less uniformly exactly 'radius' amount away from the origin. This means that they cannot group in at some random location like before and the object classes must group around them.