For long, it was assumed that the running time of a learning algorithm is directly proportional to input data; bigger the data, more is the time required to solve problems. Accepting this dogma would have a major impact on learning processes because the rate of data build-up far outpaces the growth of computational power. However, research by Volkan Cevher, Associate Professor at EPFL’s Laboratory for Information and Inference Systems, has exposed this dogma as an invalid assumption. Based on his findings, Prof. Cevher is currently engaged in ascertaining Time-Data Trade-Offs in Resource-Constrained Information and Inference Systems.
The project builds on an emerging concept in statistical computation, which considers data as a computational resource that can be leveraged to create more robust algorithms for mathematical estimation and learning. The research is based on the premise that there are trade-offs between the complexity of data and the computation time needed to arrive at accurate estimations. By implementing the trade-off, it is possible to accelerate the process of mathematical optimization.
The widespread use of machine learning and data analysis implies the use of both computational and statistical resources. In traditional studies, the resources have been studied separately. Although some researchers have adopted an interdisciplinary approach, they do not go any farther than establishing the trade-offs among the resources. On the contrary, the current research uses a robust mathematical framework to implement large-scale optimization of the time-data trade-off. That constitutes a novel approach toward the use of data in statistical sciences and could lead to much greater computational flexibility in using available resources for faster problem-solving. The study entails a series of numerical experiments and explores the geometric opportunity for a time-data trade-off.
The Time-Data project is being conducted with a Consolidator Grant from the European Research Council.