I have an assignment to implement the adaptive resonance theory (ART) type network (as part of a bigger project). I have red a lot of Internet resources on the topic and I think I've got the essence of it, but I am not sure. So far I have the following findings and questions:
- The ART1 network has two layers (input and output) which are fully connected in both directions.
- The input passes through the bottom-up connections, and the output neuron with highest value is the winner. Than the input is somehow compared against the prototype stored in the top-down connection weights (not sure how this is done, does the out-value of the winning output neuron pass through the top-down connections or not, it looks to me that the top-down connections are not really proper connections by definition)
- Does the ART1 network have a learning phase, and operating phase (like a multi-layer perceptron) or is it adjusting its weights constantly and it is up to the user to decide when is the "learning" phase over?