A Generic Dataflow Programming Environment for Neural Networks
暂无分享,去创建一个
We describe a new programming environment for neural networks (in the widest sense) based on the notion of folded dataflow graphs. Currently under development, the system will permit visual, interactive design of all aspects of neural networks (network architecture, learning algorithm, loss function, teacher/environment, etc.) within a single, generic framework, using primitives at the level of Matlab operations. The system is being implemented in C++, and is now complete except for the graphical user interface. It has been found to provide an excellent, flexible yet efficient platform for the rapid prototyping of neural networks.
[1] Barak A. Pearlmutter. Fast Exact Multiplication by the Hessian , 1994, Neural Computation.
[2] Nicol N. Schraudolph,et al. Local Gain Adaptation in Stochastic Gradient Descent , 1999 .
[3] Nicol N. Schraudolph,et al. Online Independent Component Analysis with Local Learning Rate Adaptation , 1999, NIPS.