From CNN to CNN + RNN: Adapting Visualization Techniques for Time-Series Anomaly Detection

TL;DR

Deep learning models for video anomaly detection are often opaque, raising GDPR Article $22$ concerns. The authors address this by adapting CNN visualization techniques to a CNN+RNN architecture with a TimeDistributed layer, enabling per-frame explanations via saliency maps and Grad-CAM applied to temporal data. They provide qualitative visualizations and quantitative analyses, demonstrating how per-frame attributions can accompany predictions and support transparency, while also identifying limitations inherent to time-distributed processing. The work offers a transitional, regulatory-aligned approach to explainability in video analytics and outlines concrete directions for future video-specific visualization methods and real-time deployments.

Abstract

Deep neural networks are highly effective in solving complex problems but are often viewed as "black boxes," limiting their adoption in contexts where transparency and explainability are essential. This lack of visibility raises ethical and legal concerns, particularly in critical areas like security, where automated decisions can have significant consequences. The General Data Protection Regulation (GDPR) underscores the importance of justifying these decisions. In this work, we explore visualization techniques to improve the understanding of anomaly detection models based on convolutional recurrent neural networks (CNN + RNN) with a TimeDistributed layer. Our model combines VGG19 for convolutional feature extraction and a GRU layer for sequential analysis of real-time video data. While suitable for temporal data, this structure complicates gradient propagation, as sequence elements are processed independently, dissociating temporal information. We adapt visualization techniques such as saliency maps and Grad-CAM to address these challenges. This article highlights the difficulties in visually interpreting video-based models and demonstrates how techniques for static images can be adapted to recurrent architectures, offering a transitional solution in the absence of dedicated methods.

Figures (12)

• Figure 1: Model Architecture
• Figure 2: Saliency maps
• Figure 3: Activation maps Source “tap rack gajoubert“
• Figure 4: Structure of 2D convolution / 3D convolution / Time distributed convolution
• Figure 5: Class activation maps (Grad-CAM) for a shooting video Source: "tap rack gajoubert"