Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Visual Concept Networks: A Graph-Based Approach to Detecting Anomalous Data in Deep Neural Networks

Debargha Ganguly, Debayan Gupta, Vipin Chaudhary

TL;DR

A new, straightforward method employing graph structures and topological features to effectively detect both far-OOD and near-OOD data and convert images into networks of interconnected human understandable features or visual concepts is introduced.

Abstract

Deep neural networks (DNNs), while increasingly deployed in many applications, struggle with robustness against anomalous and out-of-distribution (OOD) data. Current OOD benchmarks often oversimplify, focusing on single-object tasks and not fully representing complex real-world anomalies. This paper introduces a new, straightforward method employing graph structures and topological features to effectively detect both far-OOD and near-OOD data. We convert images into networks of interconnected human understandable features or visual concepts. Through extensive testing on two novel tasks, including ablation studies with large vocabularies and diverse tasks, we demonstrate the method's effectiveness. This approach enhances DNN resilience to OOD data and promises improved performance in various applications.

Visual Concept Networks: A Graph-Based Approach to Detecting Anomalous Data in Deep Neural Networks

TL;DR

A new, straightforward method employing graph structures and topological features to effectively detect both far-OOD and near-OOD data and convert images into networks of interconnected human understandable features or visual concepts is introduced.

Abstract

Deep neural networks (DNNs), while increasingly deployed in many applications, struggle with robustness against anomalous and out-of-distribution (OOD) data. Current OOD benchmarks often oversimplify, focusing on single-object tasks and not fully representing complex real-world anomalies. This paper introduces a new, straightforward method employing graph structures and topological features to effectively detect both far-OOD and near-OOD data. We convert images into networks of interconnected human understandable features or visual concepts. Through extensive testing on two novel tasks, including ablation studies with large vocabularies and diverse tasks, we demonstrate the method's effectiveness. This approach enhances DNN resilience to OOD data and promises improved performance in various applications.
Paper Structure (8 sections, 1 equation, 2 figures, 5 tables)

This paper contains 8 sections, 1 equation, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Problem Setup
  4. Benchmarking Tasks
  5. Graph Embedding Algorithms
  6. Evaluation Metrics
  7. Results & Discussion
  8. Conclusion

Figures (2)

  • Figure 1: An illustrative representation of the image-to-graph transformation process for out-of-distribution detection. The diagram commences with input images and their associated visual features, demarcated by bounding boxes. This visual information is then channeled into a graph structure, with nodes demonstrating unique visual elements and edges establishing interconnections. The concluding stage showcases the embedding of the entire graph into a 2D space, where similar visual patterns manifest as proximal clusters.
  • Figure 2: Far OOD evaluation metrics while using class conditioned mahalanobis score. (a) Mahalanobis distance. (b) Precision-Recall Curve. (c) ROC AUC.