Consider multiple neural networks for the same task and suppose we can optimize each network such that the output errors are as uncorrelated as possible with respect to the input data.
Does integrating the outputs of these multiple neural networks (e.g. by voting) improve the accuracy for the outputs and robustness against the attacks (e.g. adversarial examples)?
I agree with the comment of Sébastien Dourlens, if the data has no good separability then multiple nn may not work.
In general, an ensemble of neural networks should provide better results than a single neural network. You should definitely try it!
no. the complexity of the learned function depends on layers and neurons by layers number, and the learning examples sufficient to approximate and generalize but not too much to avoid memory effect.
in fact, nn can be viewed as a statistical classification method.
I agree with the comment of Sébastien Dourlens, if the data has no good separability then multiple nn may not work.
Yes off course. This is actually true for all classifiers, not only neural networks.
Review the literature on combining classifiers, and classifier fusion.
Thank you so much for many answers.
I simply anticipate independent classifiers improve the accuracy, so I asked the question to know the case of nn.
I found that there is a possiblity that multiple neural networks may not work depending on the data.
If the fusion or fission of classifiers is present, then it is not a simply a multiple nn but more an architecture with other conditional components. My answer was in the case of convenitional nn structure, not very complex structures. For example, if you want your nn to take care of the dynamics of data, you need a loop : rnn, tdnn, mmm. if you need abstraction, you will use deep nn. and so on.
- Badges
- Science topic