It is important to understand the distinction between a decision
made with low confidence andadecision made based
on low quality information. Ideally, we would like these to be
the same: when the information is poor, the decision should
be relatively ambiguous. In particular, when the information
from one modality is poor, that modality should have relatively
little in
uence on the overall decision. However, in many cases, just the opposite is true. In the case described
above | a model poorly trained for the environment in which
it is operating | low probability estimates tend to be accompanied
by inappropriately high condence. The reason
for this is that probabilities of independent observations are multiplied together. A single outlier in the observation sequence,
which has a very low probability in all states, can
exert more in
uence on the nal decision than the high probability
observations.