Answer 1: from the environment they are situated in. Say we gave a
machine a simple learning mechanism and sent it out (e.g. a web robot),
then it would learn information we did not put in, because it has access
to different information.
Answer 2: Even for self-contained purely-formal tasks this is
possible. For example I set a genetic programming algorithm that
evolves a population of purely formal programs to solve a difficult
problem (e.g. a symbolic integration). The solution it finds could be
completely new to me - I did not put the solution in to it, it emerged.
The naive (information in >= information out) equation misses the point
about the possible intricacies of intention. In this case I intended
that the system come up with a solution I could not foresee. Intention
works at different levels: one can intend (level A) the unintended
(level B).
In some abstract sense the information comes from the random-number
stream plus the initial configuration, but a shannon-type formulation of
information is not adequate here as the output has *meaning* (note that
the technique works as well with a pseudo-random number generator as a
real-world random number stream).
> If the materials that make up these machines were
> good processors of this type of knowledge, then why did biological
> lifeforms pick it up and evolve rather than resins etc.
Because some mechanisms are sufficiently delicate and brittle that the
requisite support mechanisms for it must evolve first, but after these
are available it is free to evolve (after all we *did* evolve them).
---------------------------------------------------
Bruce Edmonds,
Centre for Policy Modelling,
Manchester Metropolitan University, Aytoun Bldg.,
Aytoun St., Manchester, M1 3GH. UK.
Tel: +44 161 247 6479 Fax: +44 161 247 6802
http://bruce.edmonds.name