| Meno: | Jakub
|
|---|
| Priezvisko: | Novák
|
|---|
| Názov: | Prevention of occurrence of dead units in self-organizing maps
|
|---|
| Vedúci: | doc. RNDr. Martin Takác, PhD.
|
|---|
| Rok: | 2019
|
|---|
| Blok: | INF
|
|---|
| Kµúčové slová: | dead units, dead neurons, self-organization, self-organizing map, SOM, neural networks
|
|---|
| Abstrakt: | One of the problems of self-organizing maps is the occurrence of so-called dead units,
which cause the map to not be fully utilized. A bad initialization of weights can cause
some neurons to have their weights far from any input data, rendering them useless,
because they will never win the competition and, most likely, will not be in close
proximity to the winners neighborhood to be sufficiently adapted by the neighborhood
function.
In this thesis, we propose several modifications to the self-organizing maps training
algorithm aimed at preventing the occurrence of dead units, along with exploring their
parameter space and how the parameter values influence the outcome.
|
|---|