机器学习面临的挑战

Abstract

Machine learning and statistics are one and the same discipline, with different communities of researchers attacking essentially the same fundamental problems from different perspectives. In this note we briefly describe some current challenges in the fi eld of statistical machine learning that cut across the communities. We focus on areas where active development of learning techniques demonstrates promising performance, but where significant gaps in the theoretical foundations remain; fi lling the gaps will help to explain and improve upon this performance. The themes are high dimensional data, sparsity, semi-supervised learning, the relation between computation and risk, and structured prediction. Our selection of these themes is highly biased (and therefore has high risk), but we believe that these challenging areas can benefit from a combination of the statistics and computer science perspectives on learning from data.