Friday, April 04, 2014

Humans versus Machines, Statistics for the win.

In an online interview, Professor Chris Wiggins said
... Machine learning sits at the intersection of data engineering and mathematical modeling. The thing that makes it different from statistics traditionally, is far more focus on building algorithms.
Another difference, although this is more of a spiritual difference, is that statistics traditionally has had a stronger emphasis in explaining a data set and machine learning has far more interest in building predictive models. For example, when Netflix tells you what movie to watch or when Amazon predicts what book to buy--that’s machine learning. ...
For years, this has been my understanding of the distinction between Statistical Learning and Machine Learning but it means more when it came from a machine learning expert. Statisticians are indeed obsessed with finding explanations of the variations we observe in a data set, through modeling, visualization, simulation-based model checking and etc. We are interested in which few features can explain away a large proportion of variation in the outcome of interest, and through what forms of mechanisms. These features are mostly likely related to the scientific mechanisms behind the data. Understanding such mechanisms will be of importance to any data-driven decision making process such as policy making, intervention, and e.g., personalized medicine. We do not believe that our models are correct but we do think they are often very useful. For example, we use model-based framework to understand what assumptions were made for specific machine learning algorithms. Such understanding is then used to evaluate these algorithms under the situations where the assumptions do not hold and to suggest extensions, modifications to make them work better. Modeling is not our goal. Understanding is. Machines do not need to understand the work they are doing but humans do. For the curious minds, Statistics can help.