Abstract
Sum-product networks are a relatively new and increasingly popular class of (precise) probabilistic graphical models that allow for marginal inference with polynomial effort. As with other probabilistic models, sum-product networks are often learned from data and used to perform classification. Hence, their results are prone to be unreliable and overconfident. In this work, we develop credal sum-product networks, an imprecise extension of sum-product networks. We present algorithms and complexity results for common inference tasks. We apply our algorithms on realistic classification task using images of digits and show that credal sum-product networks obtained by a perturbation of the parameters of learned sum-product networks are able to distinguish between reliable and unreliable classifications with high accuracy.
| Original language | English |
|---|---|
| Title of host publication | ISIPTA'17: Proceedings of the Tenth International Symposium on Imprecise Probability: Theories and Applications |
| Pages | 205-216 |
| Number of pages | 12 |
| Publication status | Published - 2017 |
| Externally published | Yes |
| Event | 10th International Symposium on Imprecise Probability: Theories and Applications, ISIPTA 2017 - Lugano, Switzerland Duration: 10 Jul 2017 → 14 Jul 2017 |
Publication series
| Name | Proceedings of Machine Learning Research |
|---|
Conference
| Conference | 10th International Symposium on Imprecise Probability: Theories and Applications, ISIPTA 2017 |
|---|---|
| Country/Territory | Switzerland |
| City | Lugano |
| Period | 10/07/17 → 14/07/17 |
Funding
This work was partially supported by CNPq (grants 308433/2014-9, 303920/2016-5) and FAPESP (grants 2016/01055-1). We greatly thank Renato Geh for making his source code and the handwritten digits dataset publicly available (at http://github.com/RenatoGeh/gospn).
Keywords
- Credal classification
- Sum-product networks
- Tractable probabilistic models