Beyond Normal References: Discriminative Few-Shot Anomaly Detection

A new framework called IDEAL aims to improve few-shot anomaly detection by learning from both normal and anomalous reference examples, rather than relying solely on normal data. The approach, detailed in a paper posted to arXiv, addresses a setting the authors term discriminative FSAD [1][2]. Existing few-shot anomaly detection (FSAD) methods typically perform normality matching, comparing query samples against a small set of known normal references. This strategy ignores information contained in anomalous examples that may also be available during inference. The paper’s authors argue that directly fitting a model to both normal and anomalous references can cause it to overfit to the specific anomalies it has seen, limiting its ability to flag new types of defects [1][2]. To address this, they propose IDEAL, which stands for Intrinsic Deviation Learning. The framework is designed to extract generalizable patterns of abnormality by modeling how anomalies deviate from normality [2]. IDEAL breaks the learning process into two components. The first, a Normal Variation Eraser, suppresses irrelevant variations within normal samples that could otherwise produce noisy deviation signals. This step highlights representations that are relevant to actual anomalies. The second component, an Intrinsic Deviation Encoder, takes these cleaned-up deviation representations and decomposes them into a set of orthogonal intrinsic deviation vectors. These vectors capture the most discriminative directions of deviation [1][2]. At inference, a query sample is scored by measuring its deviation from normal references and then projecting that deviation onto the learned intrinsic deviation vectors. The preserved deviation after projection serves as the anomaly score, a mechanism the authors state enables the model to generalize to both seen and unseen anomaly types [2]. The work operates within the broader field of neural network research, where architectural innovations such as convolutional neural networks and attention mechanisms have driven progress in computer vision tasks [4]. The discriminative FSAD setting also reflects a practical reality in industrial and scientific applications: high-quality labeled datasets for supervised learning are expensive and time-consuming to produce, making techniques that can learn from very few examples particularly valuable [5]. The paper reports that experiments across eight real-world datasets showed IDEAL consistently outperformed existing state-of-the-art FSAD methods. The authors have indicated that code and data will be made publicly available on GitHub [1][2].