Neural Information Processing Systems http://nips.cc/

Our contribution can be presented in four points.

The EBP method is indeed capable of learning binary NNs in a fairly effective way. The catch is though that using binary models directly (i.e. the MAP solution) at test time directly does not seem to lead to competitive results (see column "Binary EBP-D" in Table 1: the error rate is >= doubled for 4 datasets out of 7). What seems to work well is Bayesian model averaging with the binary model ("Binary EBP-P") but this is again a'continuous' computation which is at least as expensive as the computation with a normal NN. At the very least the authors need to clarify this point, currently, as the binary version is the key motivation of this research (see the first paragraph of the paper).

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. Stochastic variational inference (SVI) requires careful selection of a step size. This paper proposes a Kalman filter to set the step size automatically. The authors show that standard Gaussian KF does not satisfy the Robbins Munro criteria (and performs badly). They propose to apply a KF based on T-distributions, and show that this gives better results than standard SVI.

Neural Information Processing Systems http://nips.cc/

The reviewer seems to have clearly understood the approach and our intended contribution. We feel that the reviewer's objections mostly involve technical issues that we did not explain or justify clearly enough, but which do not undermine the novelty or soundness of the basic results. We apologize for not explaining / justifying these issues more carefully, and feel they will be straightforward to address in the revision. We thank the reviewer for raising them.

In this paper, we introduce variational semantic memory into meta-learning to acquire long-term knowledge for few-shot learning.