Purpose of this Site
This page focuses on more practical implementations than those outlined in "On Logical Triangulation".
It focuses on my research in until the end of the year 2014, and is a kind of presentation of some of my approaches.
A multitude of example systems is presented, among them such that:
- match misspelled words,
- recognise images,
- converse with the user,
- automatically examine text (e.g. contracts),
- and learn to play a game;
- all of this using variations of one and the same algorithm.
In this theory, the ideas described in "On Logical Triangulation" are refined and simplified. It is demonstrated how systems relying on patterns (of whatever nature - sound, text, vision...) could be implemented.
Patterns, in the sense used herein, are sets of elementary atoms which are not clearly hierarchised the way that has been presented in "On Logical Triangulation", but which are handled as entities without hierarchisation.
It is shown how a symbolic artificial intelligence can interact with the outer world, based on three principles:
(a) - The artificially intelligent agent, or "system", must be able to observe the outer world's behaviour and learn from it;
(b) - the system must be able to react to the outer world;
(c) - the system must be able to handle imperfect information, i.e. it must be able to conduct approximate matches of perceptions to its experiences in order to reply sensibly.
Furthermore, it is discussed in a more philosophical perspective how the "self" is an illusion, how any individual is simultaneously a swarm intelligence, too, and what limits human consciousness really has. The conclusion is that the human is itself nothing but a pattern matching mechanism, and that therefore, an entity sophisticated enough could well match his abilities. Such an entity may well operate by means of patterns.
Finally, several ways are shown how neural networks can be improved through conclusions drawn out of pattern matching and logical triangulation. Several example systems are presented. Kindly note in particular the implementation of a multi-layer hardlimit (i.e. with step-function activation) neural network, which itself is presented as a swarm intelligence (following the above-mentioned idea that any "individual" can be represented by a swarm).
It is shown, too, how a continuously learning neural network can be implemented, i.e. one that does not have separate stages of training and usage.
Attached is the corresponding theory, in which all of the above is described, as well as an archive with the systems presented therein. Included is also the updated source code of the Katoptron as rfo20150123.bas, as it no longer corresponds exactly to the one described in the theory.
Enjoy and let me know what you think!
- Nino
k e d a l i o n [dot] d a i m o n [at-sign] g m a i l [dot] c o m
larcom-e-s.zip
onpatterns20150106.pdf
rfo20150123.bas