本文提出的是一个极其顶层的深度学习概念:Learning with angular information between w and x on Hyperspheres。本着这个精神,作者提出了SphereConve,depend on which SphereNet is established,which is then trained with Angualr Softmax loss。
SphereConv,就是一种有别于传统CNN中的卷积层的新型卷积层操作。传统卷积的activation value计算是求w(filter, or kernels, or feature template)和x(image patch or signals in general)的dot-product。而SphereConv,是projecting parameter (w) learning onto unit hyperspheres, where layer activations depend not on dot-product, but on the geodesic distance betweeen w and x. Geodesic distance, is essentially the angle between w and x.