read the Hinton paper cited in the OP -- no amount of processing power will make CNNs represent structured latent variables