Bayesian neural networks typically approximate predictions by averaging over a fixed number of Monte Carlo samples per input. That fixed budget takes no account of the approximation error introduced by using only a finite sample, so some predictions may be less reliable than others.
The paper proposes an adaptive inference scheme: rather than always drawing the same number of samples, the network continues sampling until a confidence-based criterion is satisfied. This provides statistical guarantees on the resulting predictions, addressing a gap in standard practice.
The approach matters because it links computational cost to the difficulty of each input. Easier inputs may require fewer samples, while harder ones receive more, potentially making Bayesian deep learning more efficient without sacrificing reliability.