The final categories to assign meaning and determine the usefulness of the results are adaptive resonance theory (ART) a neural network architecture that is aimed at being brain-like in unsupervised mode Kohonen self-organizing feature maps
(i.e., neural network models for machine learning)
http://www70.homepage.villanova.edu/matthew.liberatore/Mgt2206/turban_online_ch06.pdf
If data points are presented to the network one-after-another, they influence shape of the network lattice (The positions of neurons in data space). Imagine, that you have network that consist of two neurons only and train it with two data samples, that are placed on the opposite sides of the first neuron in data space. It would look somewhat like that:
A ----- neuron1 ----- B -------- neuron 2
If you start training with "A" sample and then proceed to "B", first neuron would win the competition for the "A" sample, would be dragged towards it and then loose competition with neuron 2 for the "B" sample, which would result in both neurons changing their position in the left direction.
If you reverse the order of the samples, neuron1 would win for the "B" sample, be dragged towards it then win again for the "A" sample and be dragged towards A, which would result in neurons placed in similar positions as in the beginning.
This example is very simple, yet the principle remains similar also for normal-size network and data set. I hope it answers your question.
With best regards,
Ziemowit
.
in addition to Ziemowit's excellent answer above :
this is the result of the on-line training as he describes it
if, for some reason, this (fairly small) dependence is a problem for you, you can always resort to the "batch training" but its main drawbacks are
- getting stuck more easily in local minima
- being more dependent on the weights initialization
see
https://samos.univ-paris1.fr/archives/ftp/preprints/samos158.pdf
for details
i stress that this dependence on the order for on-line training is very small as far as the final result (neuron weights) is concerned ; i have never considered it a practical problem in any way when the map is properly (slowly enough) trained
.
- Badges
- Science topic