3 Cognitive and Engineering Issues

Traditional approaches to Artificial Intelligence systems, including Computer Vision, have been highly anthropomorphic in that they have tried to emulate high-level behaviours, such as logical reasoning, symbol manipulation and identification, which are only relevant to human behaviour. The view taken here is that the key to understanding human behaviour is to first look at the foundations upon which all living systems are built and work upwards from there. Briefly raised in this section are some of the main Cognitive Science issues relevant to the modelling of AI systems and some of the engineering benefits that arise out of such a naturally parsimonious approach.

Functional Decomposition

Natural systems are more appropriately modelled as distributed systems than a collection of functionally distinct models with a central controller.

Representation

The nature of representation is probably one of the most prominent debates in the AI world. Traditional approaches have gone for models based upon explicit high-level or metric properties. The more likely candidate for representation in neural systems is of distributions of activations of neurons.

Commitment

A major goal in much vision research is to deliver reconstructed 3-D models of the world, irrespective of relevance, following the principle of least commitment. Natural animate systems, on the other hand, follow the principle of most commitment, in that it is only necessary to process what has meaning for them.

Feedback Control

Natural systems act continually upon feedback from the environment which means it is not necessary to compute specific actions. Traditional AI approaches are categorised by systems that are designed to compute specific outputs.

Situatedness

Reasoning or symbolic systems which have no connection to any environment tell us little about how systems acquire relevant semantics. Natural systems being situated in a real environment, interact with and experience meaningful events and objects directly.

Action

Active systems benefit greatly from being able to change the state of their perceived world and to acquire new information.

In consideration of producing a realistic cognitive model of a basic vision-based, animate system the policies of the non-traditional approaches and the principles of natural systems are adopted. Also noted are significant engineering benefits for systems developed along these lines.

Reconstruction

3-D mapping of an agent's environment is an extremely costly endeavour. It requires processes and techniques that can measure the world, as well as knowledge of its own position and the effects of changes to the world due to its actions. A system which behaves independently of such knowledge benefits from a simplified design that does not require calibration or measurement mechanisms.

Action

A static vision system may have to transform internal representations if its current viewpoint is not optimal. An active system, on the other hand, does not require such complex processes, but is able, by moving to a new viewpoint, to transform the world.

Perceptual Control

Continual update of its position relative to a goal perception removes the necessity to compute specific outputs, which would have lead to the architectural rationale of measure and model and its associated complexity.