SpecNN: The specifying neural network
Kirili Kochetov, Eugene Lane · 2016
In this paper we present a novel architecture design called SpecNN for artificial neural networks. Our approach allows to consider prior probability distributions and leverage samples similarity to handle the problem of fine-grained samples, thus improving the classification accuracy. We present two different learning algorithms for SpecNN. SpecNN is especially useful in moot cases, when classification is instable. Experiments conducted on several datasets, including MNIST, show that SpecNN outperforms multilayer perceptron.