Adaptive Cluster Expansion (ACE): A Hierarchical Bayesian Network

TL;DR

The paper introduces Adaptive Cluster Expansion (ACE), a hierarchical Bayesian network that efficiently estimates high-dimensional probability densities and detects anomalies in images using hierarchical vector quantization.

Contribution

It presents the ACE method, which captures high-order statistics quickly through hierarchical vector quantization, and introduces a scheme for visualizing network states as probability images.

Findings

ACE effectively captures high-order statistics with short training times.

ACE can identify subtle anomalies in texture images.

Probability images reveal statistically anomalous regions.

Abstract

Using the maximum entropy method, we derive the "adaptive cluster expansion" (ACE), which can be trained to estimate probability density functions in high dimensional spaces. The main advantage of ACE over other Bayesian networks is its ability to capture high order statistics after short training times, which it achieves by making use of a hierarchical vector quantisation of the input data. We derive a scheme for representing the state of an ACE network as a "probability image", which allows us to identify statistically anomalous regions in an otherwise statistically homogeneous image, for instance. Finally, we present some probability images that we obtained after training ACE on some Brodatz texture images - these demonstrate the ability of ACE to detect subtle textural anomalies.