Python Programming for Biology: Bioinformatics and Beyond

Supervised machine learning

The  supervised  kind  of  machine  learning  involves  having  a  computer  algorithm  that  we

can train on some known data. Here you would have some input data and knowledge of

what  each  piece  of  data  corresponds  to.  Using  the  postal  code  handwriting  example,  the

input data might be examples of handwritten marks and the output would be knowledge of

which of the 10 numerals or 26 letters of the alphabet were written. To take a biological

example  you  may  have  a  set  of  DNA  or  protein  sequences  and  associate  each  with  an

experimentally  determined  category.  Initially,  during  the  training  stage,  our  special

computer  algorithm  learns  to  associate  each  input  with  the  correct,  known,  output  by

adapting: essentially changing some internal parameters so that if you present an input to

the  algorithm  it  gives  an  output  that  is  as  close  to  the  known  result  as  possible.  The

objective  is  usually  to  have  a  single  set  of  parameters  that  performs  the  job  in  a  general

way, adapting to all of the data to learn the overall properties of the problem, rather than

optimising  performance  for  some  data  at  the  expense  of  others.  In  this  way,  after  the

training stage the parameters of the algorithm are fixed and the algorithm may be applied

to  input  data  it  has  not  seen  before  and  come  up  with  an  output  prediction:  a  prediction

made following the rules learned during the training phase.

Broadly  speaking  you  will  encounter  supervised  machine  learning  being  used  in  two

major ways: to generate discrete alternative outputs, for example to perform classification

as with the reading of handwritten letters and numbers; and also to predict the value of a

continuous  variable,  which  is  often  referred  to  as  functional  approximation.  Predicting  a

continuous  variable  would  occur  when  you  are  predicting  something  on  a  sliding  scale,

like temperature, energy etc. It is perhaps fortunate that we can often use the same basic

kinds  of  computational  algorithm,  albeit  with  a  degree  of  modification,  whether  we  are

predicting a discrete classification or a continuous variable. In the case of artificial neural

network  methods,  which  we  introduce  below,  the  distinction  between  the  two  types  of

situation can be as simple as restricting values to 0.0 or 1.0 if the problem is of the discrete

kind.

---