Brain-eating as an efficient way of learning

More than 4 months have passed since the launch of Qubic. During these months 100s computers were mining and at the same time producing information for Aigarth, a project running on top of Qubic. Billions artificial neural networks (ANNs) were created and destroyed and this process revealed an interesting thing.

It appears that ANNs containing fragments of those ANNs which had a low value of the fitness function and were hence “killed” can be trained significantly faster.

[So, learning how to “kill” other ANNs and salvaging their dead “bodies” is an efficient way of surviving (because dumb ANNs are “killed” by us during the training.) If we recall our history we’ll notice that we were killing each other all the time since many thousand years back. And we became much smarter during that. Maybe a coincidence, maybe not…]

Anyway, let’s return to our ANNs. If ANNs made of fragments of less fit ANNs are trained faster then can we save a lot of resources if we generate (without any training) already “dead” ANNs and “feed” yet alive ANNs with them? This is something that we are trying to find out with mining happening right now.