Scientific Presuppositions

 

Some of the premises of logical AI are scientific in the sense that they are subject to scientific verification. This may also be true of some of the premises listed above as philosophical.

 

innate knowledge

The human brain has important innate knowledge, e.g. that the world includes three dimensional objects that usually persist even when not observed. This was learned by evolution. Acquiring such knowledge by learning from sense data will be quite hard. It is better to build it into AI systems.

Different animals have different innate knowledge. Dogs know about permanent objects and will look for them when they are hidden. Very likely, cockroaches don't know about objects.

Identifying human innate knowledge has been the subject of recent psychological research. See [Spelke 1994] and the discussion in [Pinker 1997] and the references Pinker gives. In particular, babies and dogs know innately that there are permanent objects and look for them when they go out of sight. We'd better build that in.

middle out

Humans deal with middle-sized objects and develop our knowledge up and down from the middle. Formal theories of the world must also start from the middle where our experience informs us. Efforts to start from the most basic concepts, e.g. to make a basic ontology are unlikely to succeed as well as starting in the middle. The ontology must be compatible with the idea that the basic entities in the ontology are not the basic entities in the world. More basic entities are known less well than the middle entities.

universality of intelligence

Achieving goals in the world requires that an agent with limited knowledge, computational ability and ability to observe use certain methods. This is independent of whether the agent is human, Martian or machine. For example, playing chess-like games effectively requires something like alpha-beta pruning. Perhaps this should be regarded as a scientific opinion (or bet) rather than as philosophical.

 

universal expressiveness of logic

This is a proposition analogous to the Turing thesis that Turing machines are computationally universal--anything that can be computed by any machine can be computed by a Turing machine. The expressiveness thesis is that anything that can be expressed, can be expressed in first order logic. Some elaboration of the idea is required before it will be as clear as the Turing thesis.

 

sufficient complexity yields essentially unique interpretations

A robot that interacts with the world in a sufficiently complex way gives rise to an essentially unique interpretation of the part of the world with which it interacts. This is an empirical, scientific proposition, but many people, especially philosophers (see [Quine 1969], [Putnam 1975], [Dennett 1971], [Dennett 1998]), take its negation for granted. There are often many interpretations in the world of short descriptions, but long descriptions almost always admit at most one.

The most straightforward example is that a simple substitution cipher cryptogram of an English sentence usually has multiple interpretations if the text is less than 21 letters and usually has a unique interpretation if the text is longer than 21 letters. Why 21? It's a measure of the redundancy of English. The redundancy of a person's or a robot's interaction with the world is just as real--though clearly much harder to quantify.

We expect these philosophical and scientific presuppositions to become more important as AI begins to tackle human level intelligence.