Re: Impressively Smart, Non-Evolved Neural Network
- Hello everybody,
with regard to Spaun I agree with the others that it is an interesting project in its own right but does not really compete with Neuroevolution.
> More generally, I'd be interested in any of your reactions regarding whether you agree that this is a very impressive level of intelligenceI disagree that this is impressive when intelligence is the measure. You can recognize that in the paper itself. For example, what working memory is and how it works is part of ongoing research and we are far away from understanding it. So one could guess that Spaun could help us to better understand how working memory works. But what becomes apparent in their paper is that they 'just' implemented their own model of working memory. Thus, no further insights about this cognitive ability can be gained.
> and whether that means we should question our devotion to neuroevolution as one of the quickest paths to AGI.I totally agree with that. I can perfectly live with the fact that hand design can be in some cases faster than a neuroevolution approach. But I think, we all are interested in neuroevolution in order to learn something about the basic principles nature uses to build complex network and thereby complex behavior?!
Although I find it quite hard to incorporate the methods that have been developed in this field (e.g. NEAT and CPPNS) into my studies that aim to understand natural brain by building artificial models, I think in the long run only an unbiased search through topologies can help us to find the biological concepts through which complex behavior can be implemented. I think, we still understand too less about the brain in order to build it in an objective driven approach. Instead, an evolutionary approach could even constrain the search space in order to find scalable and plastic networks that can handle complex tasks. Admittedly, developing the right methods to do that is a difficult task on its own.