Design human-like machine learning

A study of the publications available via PubMed alone (a database of 26 million references to biomedical publications), shows that only four machine learning models (linear regression, Bayesian reasoning, Random Forests and Neural Networks) are frequently employed in most of the experiments. Non-experienced machine learning users tend to always use some standard methods which have previously been used in similar studies. For a non-expert, it is difficult to know which method to use when. Even a 5% error difference may have a huge impact in some cases such as prescribing a drug, for instance. The models themselves have their limitations and there is a whole branch of computer scientists who work on theoretical aspects and prove the bounds of these models.
A solution to this unnecessary scientific uncertainty will be automated machine learning. This automated machine learning should be able to imitate all aspects of human information acquisition: (i) learn (from experience or from data); (ii) unlearn (or forget information that is no longer valid); (iii) relearn (when previous information has not been corrected acquired); and (iv) be taught (via an optimal number of examples). There isn’t any complex methodology available which includes all these building blocks together and makes use of them. This will be the aim of this project, resulting in an easy to access framework for advanced data analysis and visualisation; an intelligent assistant which could interact with the user, could self-learn from experience based on similarity between the data analysed so far and which - based on a research on (already existing) research - could ensure the quality of the results through repetitive tests, automating the automation of knowledge discovery.

A strong computer science (machine learning) background is required as well as strong programming skills (Python, R, Java, C++).