> Other colleagues maintain that we should be able to construct large, uniform neural networks that can learn to do all that minds might need. I do not see

> Other colleagues maintain that we should be able to construct large, uniform neural networks that can learn to do all that minds might need. I do not see much hope of this, because of fear that any very large such network would be prone to accumulate too many interconnections and become paralyzed by oscillations or instabilities. How could we stabilize such systems ? My answer is that one might have to provide a variety of alternative sub-systems, decoupled enough that if each part should fail from time to time, the rest could continue to function so that not all the system will all fail at once. This means that those parts must be suitably insulated from one another. Looks like Generative Adversorial Networks ?