Novelty of learning rule and key contributions (R2): While the final expression for our learning rule is superficially

We thank the reviewers (R1, R2, R3) for their detailed feedback. Key concerns are briefly addressed below. Kolen & Pollack, 1994), so symmetry is an emergent property of the network. Prior manipulations were made throughout learning, not only at test time. "It is not clear in what way intrinsic noise in the recurrent dynamics ... allows us to derive closed-form probabilistic ": when we say that noise is necessary for our learning rule, we mean Intrinsic noise is what allows for this probabilistic description.