Abstract
We propose an efficient procedure for significance determination in high-dimensional dependence learning based on surrogate data testing, termed inverse finite-size scaling (IFSS). The IFSS method is based on our discovery of a universal scaling property of random matrices which enables inference about signal behavior from much smaller scale surrogate data than the dimensionality of the original data. As a motivating example, we demonstrate the procedure for ultra-high-dimensional Potts models with order of 1010 parameters. IFSS reduces the computational effort of the data-testing procedure by several orders of magnitude, making it very efficient for practical purposes. This approach thus holds considerable potential for generalization to other types of complex models.
Original language | English |
---|---|
Article number | 062112 |
Pages (from-to) | 1-9 |
Journal | Physical Review E |
Volume | 97 |
Issue number | 6 |
DOIs | |
Publication status | Published - 6 Jun 2018 |
MoE publication type | A1 Journal article-refereed |