Machine Learning with Distances
Abstract
Various machine learning tasks such as learning under non-stationarity, change detection, and dimensionality reduction can be solved by estimating some distance/ratio between two probability distributions. This project developed accurate and computationally efficient methods for estimating the distance/ratio from data, and demonstrated their usefulness in experiments. The principle idea is that when solving a problem of interest, we should not solve a more general sub-problem as an intermediate step, i,e., directly estimate the difference/ratio of the two distributions rather than estimating both separately and take the difference/ratio later. Types of the problems actually solved are change detection in time series, salient object detection in an image, measuring statistical independence, detection of structure change, covariance shift, class balance change, information maximization clustering.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 16, 2015
- Accession Number
- ADA616936
Entities
People
- Masashi Sugiyama
Organizations
- Tokyo Institute of Technology