2008
The following papers are concerned with doing inference on very large datasets that do not fit in memory all at once. Incremental learning algorithms are derived for nonparametric Bayesian models. The algorithms are based on mean-field variational approximate inference, and they alternate between "model building" and "model compression" stages. The model complexity (e.g. number of clusters) adapts as necessary as more data is processed.
- R. Gomes, M. Welling, and P. Perona (2008). Memory bounded inference in topic models. Proceedings of the International Conference of Machine Learning.
Paper | Bibtex | Video Lecture | Slides
Note: Typos were corrected in this version of the paper. In a number of places the sufficient statistics \phi_{l}(x) were written instead as \phi_{kl}(x). (There is no dependence on k). - R. Gomes, M. Welling, and P. Perona (2008). Incremental learning of nonparametric Bayesian mixture models. Proceedings of Computer Vision and Pattern Recognition.
Paper | Bibtex