At first blush, this is a terrible idea.  Try to Monte Carlo sample weights of any decent sized neural network on a GPU.  Get back to me when you’re done running that…  this is far from efficient on today’s systems. There is a reason most sampling on things like VAEs and dropout is on neurons. 2/

On the other hand, neuromorphic architectures don’t have the same constraints. So why can’t we stick a capability like sampling of weights into there as well? In particular, in our COINFLIPS project we’re examining devices like magnetic tunnel junctions that can be a weighted coin.  3/