Goto

Collaborating Authors

 mfvi and mcdo


Supplement: OntheExpressivenessofApproximate InferenceinBayesianNeuralNetworks

Neural Information Processing Systems

For each plot, we show the predictive mean and2 standard deviations along theline segment in input spacejoining the centres of these two clusters. Consider a single-hidden layer ReLU neural network mapping from RD RK with I N hidden units. Each row represents the same random dataset. Note that the data appears noisy, but this is due to the projection onto a lowerdimensionalspace. N(vi;µvi,σ2vi), (1) where wi = {wk,i}Kk=1 are the weights out of neuron i and b = {bk}Kk=1 are the output biases, and qi(wi|U,v) and q(b|U,v) are arbitrary probability densities with finite first two moments.