Fig. 2. Illustration of the problem of reliability in deep learning.

The illustration depicts the use of a neural network to predict the anomalous exponent α for two sample trajectories. Despite receiving severely different inputs, a classical neural network may still predict the same output (anomalous diffusion exponent α = 1) for both cases. The difference between the outputs only becomes clear when predicting not just the output itself but a distribution over all possible outputs, as it is done, for example, in Bayesian Deep Learning.