2010_Deep big simple neural nets excel on hand-written digit recognition
首先看这篇文章:
2003_Best practice for convolutional neural networks applied to visual document analysis这篇文章中心思想:
Expanding training set with elastic deformation and affine transformation
本文的中心思想是运用仿射变换和形变扩大数据集,然后运用GPU加速训练庞大的简单神经网络,从而获得比较好的结果。
实验结果:
结论:
1. convolutional network is better than MLP
2. Elastic deformations have a considerable impact on performance both for 2-layer MLP and convolutional network
3. Cross-entropy seems better and faster than mean square error(MSE)
https://blog.sciencenet.cn/blog-1583812-814176.html
上一篇:
An analysis of single-layer network in unsupervised...下一篇:
1998_Efficient Backprop笔记