1 Erratum to: Neural Process Lett DOI 10.1007/s11063-015-9444-3

The original version of this article unfortunately contained a mistake. The presentation of Fig. 1a, b was incorrect. The corrected version is given below.

Fig. 1
figure 1

a The formulation in [23]: a data point \(\varvec{x} \in {\mathcal {X}}\) is mapped into m feature spaces via \(\phi _1,\phi _2,\ldots ,\phi _M\), which are then scaled by \(\mu _1,\mu _2,\ldots ,\mu _M\) to form a weighted feature space \({\mathcal {H}}^*\), which is subsequently projected to the embedding space via an universal projection \(\varvec{L}\). b The formulation in [12]: \(\varvec{x}\) is first mapped into each kernel’s feature space and then its image in each space is directly projected into an Euclidean space via the corresponding projections \(\varvec{L}_1, \varvec{L}_2, \ldots , \varvec{L}_M\), thus the embedding space can be seen as an unweighted concatenation of the M projected Euclidean spaces. c Our proposed formulation is the weighted version of b, the projections and the weights are jointly learned to produce the embedding space, a weighted combination, b concatenated projection, c our formulation