Feature-based detection techniques have been advocated for robust spectrum sensing in cognitive radios. Cognitive radios must be able to train themselves to identify the features for a specific primary user at a given channel, time or location. However, 'in-the-field' training relies on signal observations where there is uncertainty about whether or not it is truly representative of the primary user. This work considers this uncertainty, how it effects the detector's training time and performance, and identifies a trade-off between these outcomes. A two-stage detector structure is also illustrated to fulfill both the training and operational requirements of such detectors.