Prediction Based Analysis in Gene Expression Experiments Sample Clauses

Prediction Based Analysis in Gene Expression Experiments. In biology, mechanisms play an important role. As we saw in the genetics Sec- tion (1.2), biology has developed very complex strategies during evolution. As evolution builds upon old strategies [21] there is a lot of redundancy associated with that complexity. This complexity is not easily fully described and makes tak- ing all mechanisms into account a very hard task. It is useful to model the real biology by more general, but complex enough machine learning techniques on a predictive bases. Machine learning techniques learn from seeing data, adjusting the (hyper-)parameters to fit a general view. This general view, usually does not include underlying biological mechanisms. Prediction based analysis is essentially a “black box” based approach. The ma- chine learning designer design the underlying functional relations inside this black box. They then supply intuitional insights into the function of the black box for other researchers (preferably crossing expertise), so that they can apply the algo- rithm to their problem. It is important to point out, that the design of the internals of the black box can have big impacts on results and should not be disregarded completely. However, researchers from other fields of expertise should be able to apply such complex algorithms to their own data without having to fully know the exact underlying code and technique. You should essentially be able to provide the training inputs and output pairs and let the machine learn the patterns in the data (by fitting parameters). After the learning period, we can ask the black box for likely outcomes of newly seen inputs. 4.2.1). In summary, machine learning is to restrain from direct mechanistic modelling of expected patterns and go closer to giving the machine enough resources to gener- alize over patterns itself and decide how to use the patterns in the inputs to extrap- olate on the outputs. This can significantly improve results, as long as the provided mechanisms are general enough to generalize over patterns, while not being too general to overfit the training data, over explaining patterns directly.