6373Re: DDFA Questions
- Jun 19, 2014Hi Andy, here are answers to your two recent questions:
1) During feature discovery it's probably easiest not to think about "the network" but rather the "set of networks," where each network is a single feature. It is true that eventually a bunch of these networks will be combined into a single classifier network for the final classification task, but during feature discovery that final network does not yet need to be considered. Therefore, focusing only on the feature networks, each one includes only a 28x28 input layer (the size of an MNIST digit) and a single output node (the feature detector output for that particular feature). In this experiment in this paper, these individual feature detector networks do not have hidden nodes, so they are just direct input->output one-layer networks. Of course in principle they could have been evolved with multiple layers, but our aim in the introductory paper was to demonstrate the concept in the simplest possible implementation.
2) I believe your characterization of the whole classification network is accurate. It has the 28x28 input layer, then the set of collected feature detectors in the hidden layer, and then 10 outputs (one for each possible digit). The whole thing is indeed trained with labelled examples to associate the right output with each input example. Later testing of course is with a separate set of data.
- << Previous post in topic