Towards more efficient machine learning algorithms: theory and practice

点击翻译按钮获取中文摘要

Machine learning is concerned with the development of intelligent computer systems that are able to learn and generalize from collected data. This ability is required for many important tasks that are just too complex to be explicitly programmed by humans such as recognizing voice directives on smart phones, or recognizing faces in images, or displaying the best results from a query. The best computer systems that achieve these tasks have in common the fact that their ability has been acquired by running learning algorithms on vast amounts of data. Many of these tasks are now so vital to our economy that machine learning-based technology is now commonplace. However, that technology needs to be greatly improved. Indeed, credit cards are too frequently being blocked, we are still receiving too many undesirable emails, and automatic speech recognition is still not satisfactory. Building more powerful hardware is just part of the solution as we must also find the most efficient learning algorithms. Consequently, the proposed research program aims at addressing the important problem of how to improve and modify existing learning algorithms such that they yield better predictors while using a minimal (or, at least, an acceptable) amount of resources. To meet this objective we plan to approach challenging machine-learning problems from both a theoretical and a practical perspective. Theoretical analysis is needed because we want to find efficient learning algorithms with provable guarantees. More specifically, we plan, in the short term, to improve the existing learning algorithms for structured output prediction and domain adaptation. These are currently two challenging machine-learning problems that model situations often encountered in practice but for which we do not have yet satisfactory learning algorithms. We also plan to find the most efficient boosting-type algorithms for learning from large-scale data sets. Finally, in the long term, we plan to expand the set covering machine and the decision list machine for predicting phenotypes from genomic data. These are learning algorithms (that we have proposed in the past) that produce uncharacteristically sparse predictors that are easy to interpret. Hence, these predictors can, in principle, help us to uncover the genomic cause of a phenotype. The impressive results we have obtained recently on the prediction of antibiotic resistance from bacterial genomes motivate us to consider, in the short term, the more ambitious problem of predicting cancer types from human sequences, which could be DNA, RNA, or protein sequences. It is thus expected that this research program will yield provably efficient learning methods for different machine learning-based applications that, in turn, will improve our quality of life. Moreover, it is expected that this research program will deliver the equivalent of two M.Sc graduates and four Ph.D graduates. **