We have introduced a model to analyze how a visual system can adapt to its environment, and start detecting visual events using a minimal amount of prior information. The point of view is that, although the system is hard wired (with a prescribed connectivity), it has to decide alone which visual events are relevant to describe its environment, based on some predefined organizing principles.

These principles, close to Barlow's notion of suspicious coincidences, define notable visual events as atypical complex combinations of cues. From an information theoretic point of view, they correspond to events that the system cannot encode efficiently. This implies that the system has its own notion of typical events and is able to detect the untypical ones.

This is implemented as follows. The system first extracts a series of ternary layers from an observed image, (above) based on low-level thresholded filtering.

The statistical distribution of these layers is then estimated over a large image database (which models the visual environment) under the form of a ternary Markov random field. When a new image is presented to the system (below), a statistical test is performed on local patches to decide whether a signal should be triggered.