Aligned deep learning as a random sampling method

Leonardo Pedro

Abstract: We show that the alignment problem in deep learning can be separated in two: A) estimate systematic uncertainties as in Engineering; B) new due to Big Data, how to randomly sample computable models from a set of Bayesian models, almost all uncomputable. Deep neural nets solve B). Since in most sufficiently complex systems, it is unlikely to find a sufficiently simple statistical metric that it is not misleading (or rogue, for instance the GDP in economy), the (dis)alignment is not a problem but a feature: if the estimate A) would not be likely misleading, then the sampling B) would not be random/unpredictable and if it would not be random it would not generalize well (by definition of generalization in a Bayesian context).