MOCCA: Multilayer One-Class Classification for Anomaly Detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Anomalies are frequent in scientific data, often caused by unknown processes or incomplete data knowledge.
• Training deep learning (DL) models for anomaly detection (AD) typically uses only non-anomalous data due to the rarity of anomalies.
• Existing methods often treat neural networks as single blocks, neglecting their multilayer structure.

Purpose of the Study:

• To introduce a novel framework, multilayer one-class classification (MOCCA), for training and testing DL models on the AD task.
• To explicitly optimize intermediate representations within deep architectures for anomaly detection.
• To enhance anomaly detection performance by leveraging the full multilayer structure of deep networks.

Main Methods:

• MOCCA applies a two-step training process to autoencoders for anomaly detection.
• The first step involves training the autoencoder on a standard reconstruction task.
• The second step retains the encoder and optimizes each layer's feature space by minimizing L2 distance to a data centroid, then combines features for inference.

Main Results:

• MOCCA was evaluated on CIFAR10, MVTec AD, and ShanghaiTech datasets.
• Models trained with MOCCA achieved performance comparable or superior to existing state-of-the-art methods.
• Extensive experiments and model analysis demonstrated the benefits of the proposed training procedure.

Conclusions:

• MOCCA offers an effective approach to anomaly detection by utilizing the multilayer architecture of deep learning models.
• Explicitly optimizing intermediate representations significantly improves AD performance.
• The framework provides valuable insights into the training dynamics and benefits of deep feature utilization for anomaly detection.