Deep Learning Reflection _ deep learning

Deep learning reflection with the improvement of computer hardware performance, in-depth learning in today's era as the darling, Computer vision,data mining,nature Language Process .... All take the deep learning of the car, and finally sat on the Boeing airliner. One after another, the framework of depth learning emerges, caffe,tensor flow,cxxnet, mxnet ..., and these frameworks have been disdained to blend in with single card loops, using Hadoop and the fuel that spark these big data operations, It's like building a rocket. These in-depth study of the rich do not put the cost and scale of the eye, which is more like the rich second generation, haha. The title of this article is to give these rich second-generation send some perfunctory warning. The model of the final fitting of the depth study, the classification of the data is indeed very effective, and the boundaries of the data can really be divided very well. These training models are equivalent to the result of the superposition of many nonlinear functions. Few paper pay attention to the relativity of these non-linear functions, if we can determine the connection between the neurons and weights, the lines of each neuron, and determine the space composed of the nonlinear functions of these neurons, and extract the orthogonal function set, can reduce the size of the network. The future is to study this direction, but it is true "humble" see, I hope we do not laugh