Rethinking Assumptions in Deep Anomaly Detection