Neural network training time?

Question

Neural networks

Recently, there have been a lot of articles on Habré about neural networks. Question. How much time, data, and what computing power does a neural network need to, for example, learn to identify dogs in pictures?
Are the current neural networks closed projects or can you download such a network on jithub and run it somewhere?

Answer 1

The training time of neural networks varies greatly depending on the size of the network and the size of the training sample. Of course, if you want to build a network that will produce results comparable to the winners on ImageNet, you need a comparable architecture of hundreds or thousands of machines.
But to get acquainted with neural networks and test your strength in a small but interesting application, an ordinary desktop with a discrete graphics card of not too ancient years is quite enough (for CUDA to work). There are many open libraries for this - Caffe, Theano, Torch. They have tutorials and samples with ready-made networks on which you can experiment. Recently, another excellent GUI system for training has appeared - NVIDIA DiGiTs. In general, it can be used with a minimum of special knowledge.
For example, a small Convolutional Neural Net for OCR, including numbers, letters, punctuation, etc. learns in a couple of minutes, photo analysis with several classes (nature vs city, for example) in 10-20 minutes. The most difficult thing is to set a competent internal network structure, but for a start you can use the sample ones. Well, to collect data, but everything is in your hands.

Answer 2

NS training is an art form. The result completely depends on the input data, the method of their preprocessing, and even on the sequence in which you present the input images. Each time you can get a completely different result.

Answer 3

To minimize errors, huge computing resources are needed.
Google's neural network has learned to detect cats
As for the source code, you are unlikely to find anything worthwhile on the Internet. In open source, however, there are many libraries for working with neural networks and you can build your own if you wish.