An introduction to NSN's
 








Smart Patterns

Basic concept

This applet builds a network pattern, based on morphological similarities between words.

The original network is multidimensional, but for graphic purposes, it has been translated into a 2D system : this explains why we can obtain multiple apparent stable states in some situations. All values for the construction of the network are injected through a morphological analyzer.

So, is this applet really smart ? It may well look smart, but the fact is that it is not really intelligent ! Why ? Because it lacks the main thing that makes a network intelligent : the learning skills.

Actually, this applet uses a fixed - let's say neutral - methodology to perform word pattern matching. The methodology is made out of a set of clues (indicators) that are to be balanced with each other. But like in a police investigation, the clues are usually not equal : they don't carry the same weight. The main problem in a pattern analysis process (whether human- or machine-run) is to determine the relative weights of the different available clues, in order to produce the best judgment possible in a particular situation.


Neurosemantic networks

NSN's - neurosemantic networks - were originally designed to assign weights to such clues. These networks only have a few resemblances to their fellow neural networks : a learning process (different from backpropagation), neurons, weights assigned to neuron connections, an input layer, an inner layer, and an output layer.

NSN's are complex stuff, and usually act as strategy generators. Unlike traditional artificial neural networks, their input layer will never accept direct input of values from the outside world. Instead, it will only accept objects (which can be any kind of data), and will generate internal flows, provided that specific sensors exist for those objects.

Their output layer is directly connected to motor organs, which use the output flows to build new objects from scratch, or to modify existing ones (continuous processes).

The pattern matching methodology used in the Smart Patterns applet is an object which was designed by such a NSN, after a short learning process. The learning process consisted in the NSN to propose outputs to a human operator, whose job was to answer something like "good boy, you're doing all right", or "how can you be so stupid ?". The purpose of the learning process is to make the NSN react to its teacher's answers, by adapting its own weights, and upgrading its methodology output.

NSN technologies are now at an experimental stage, and are mainly used in semantic analysis, e.g. to merge databases having similar semantic content, but different encoding.


Other applets


Copyright © Neural Semantics sprl