How Far Can Untrained Neural Networks Go in Anomaly Detection? From Tables to Pixels A study examining untrained neural networks for anomaly detection finds that intrinsic separability of anomalies from normal references can drive performance without learned features or AD-specific optimization, challenging the view that strong results require representation learning. The research spans both tabular and pixel data, according to the paper's framing. Anomaly detection AD is often framed as a representation-learning problem, where strong performance is expected to require learned features or AD-specific optimization. However, this view can overlook the intrinsic separability of anomalies from normal references, even in the absence of learned re