machine learning - Estimating the number of neurons and number of layers of an artificial neural network -
I am thinking about the number of layers and how to calculate the number of per-layer neurons. As input, I have the size of the input vector, the size of the output vector and the size of the training set.
Generally the best nets trying different net topologies and one less error, unfortunately I can not do this.
This is a very difficult problem.
The more internal structure is a network, the better it will be to represent the complex complex solution. On the other hand, too much internal structure may be slow, training can vary, or overfitting - which will prevent your network from normalizing to new data.
People have traditionally contacted this problem in various ways:
Try different configurations, see what happens best. You can divide your training into two pieces - one for training, evaluation for one - and then train and evaluate different approaches Unfortunately, it seems like your case that this experimental approach is not available. .
Use thumb rule. Many people have come as to what works best, in relation to the number of neurons in the hidden layer, people have estimated that (for example) this (A) input and output layer Should be between the size, some (inputs + output) near the (B) * is set to 2/3, or (C) never larger than the size of the input layer twice. The problem with thumb rules is that they always do not take into account the important pieces of information , and so on. Consequently, the use of these rules often starts roughly "Let's try-points-one-bunch-things-for-view-what-best-best" approach points.
Use an algorithm dynamically adjusts the network configuration. starts with the minimum network, such as the algorithm, then adding hidden nodes during training can make your experimental setup a bit easier, and (in theory) may be the result of better performance (because you accidentally hid Will not use inappropriate numbers of nodes).
There are a lot of research on this subject - so if you are really interested, there is a lot to read. See quotes, especially:
-
Lawrence, S., Giles, C.L., and Tsooi, A.C. (1996). Technical Report UMIACS-TR-96-22 and CS-TR-3617, Institute of Advanced Computer Studies, University of Maryland, College Park.
-
Elisef, A., and Pegm-Moisei, H. (1997). Advance in Neural Information Processing System 9, Cambridge, MA: MIT Press, pp.162-168.
Comments
Post a Comment