Member-only story
Physics-based AI
Why are we not interested in physics-based ANNs? Would that not be a smart way to endow a model with prior knowledge?
While it may be useful for engineering and even some robotics applications, it is not an AI approach. It does not build or use intelligence. Physics has to be taught, it is a highly abstract world-view that is built on a more direct, intuitive and simpler world model, sometimes called “naive physics”. This constitutes a system of categories which develops, instead or being learned. It requires a natural environment (for a human) plus the maturation of the brain. While it is linked to language, it is not specific to any particular language, or more precisely, small differences based on language identity become minor variations on a shared world model (e.g. color perception is marginally influenced by color naming; but the basic color terms are considerably stable across languages). It is therefore natural intelligence as the basis of ‘computational models of intelligence’ = artificial intelligence. Artificial intelligence is not an engineering discipline. It is by definition something else, a mixture of natural science, humanity and engineering, where not just any solution for a problem counts, instead a whole set of human-like abilities are developed, which may then be used and refined to solve problems, including engineering problems.
Of course we need not just models of how intelligence develops and matures, but also of how specific abstract theories can be taught and learned. Bringing abstract physics theory into a computational system, however, remains an engineering problem. It does not contribute to the specific goals of artificial intelligence with their orientation towards human abilities. We need to start more modestly.
