Python Programming for Biology: Bioinformatics and Beyond

Self-organising maps

The next machine-learning example is an illustration of unsupervised learning, where we

don’t  pass  input  data  with  any  associated  known  classification  or  value.  Instead,  the

algorithm will simply organise the data, in this case putting similar points of data near one

another and separating dissimilar points. The example we give is called a self-organising

map, and this form is also known as a Kohonen map.

5

The  reason  for  organising  data  in

this way naturally depends upon the kind of problem being addressed. Once the data has

been  rearranged  into  an  organised  form,  the  ‘map’  of  the  data,  it  can  be  divided  into

different  regions  to  define  categories.  Also,  because  the  organising  mechanism  may

consider the similarity of large numbers of features from the input data (large vectors) but

creates  a  low-dimensionality  (often  two-dimensional)  map,  the  organisation  process

performs  a  kind  of  dimensional  reduction;  we  can  more  easily  visualise  the  major

differences and similarities in the data without having to think in n-dimensional space.

The  self-organising  map  presented  here  is  the  first  example  of  an  artificial  neural

network.  Although  this  name  originally  stems  from  analogies  to  how  brain  cells  interact

with  one  another,  in  the  computational  sense  a  neural  network  may  be  imagined  as  a

network  of  interconnected  data  points  or  nodes.  Each  node  may  be  connected  to  several

others and the strength of the connections is determined by a weighting. What a data node

means  depends  somewhat  on  the  kind  of  neural  network  being  used.  For  the  self-

organising  map  example  each  node  will  represent  a  position  on  a  rectangular

6

 grid  that

makes up the map dimensions. Each grid node will possess a feature vector that is of the

same  kind  as  the  input  data  (in  the  test  example  this  will  be  a  colour).  The  features  of

different  nodes  will  be  moved  towards  and  away  from  the  input  data  vectors  so  that

different kinds of input are mapped to different spots on the grid. Here it should be noted

that the strength of the connections between the grid nodes does not change; effectively a

node will always stay in the same grid position, but how each node maps to the input data

varies.  For  the  feed-forward  neural  network  described  later  the  strength  of  connections

between nodes does change.

---