Evolution has determined a good architecture. The weight training is then just the final tweak to get everything running smoothly. No reason beyond compute we

Evolution has determined a good architecture. The weight training is then just the final tweak to get everything running smoothly. No reason beyond compute we couldn't do something similar. Ie find good architectures by evaluating them using multiple random weights, and evolve those archigectures that on average gives the best results. Then over time add a short training step before evaluating.