File Download
There are no files associated with this item.
Supplementary
-
Citations:
- Appears in Collections:
Conference Paper: Entropy coding for training deep belief networks with imbalanced and unlabeled data
Title | Entropy coding for training deep belief networks with imbalanced and unlabeled data |
---|---|
Authors | |
Issue Date | 2011 |
Citation | The 28th International Conference on Machine Learning (ICML 2011), Bellevue, WA., 28 June-2 July 2011. How to Cite? |
Abstract | Training deep belief networks (DBNs) is nor- mally done with large data sets. In this work, our goal is to predict traces of the surface of the tongue in ultrasound images of the mouth during speech. Performance on this task can be dramatically enhanced by pre- training a DBN jointly on human-supplied traces and ultrasound images, then training a modified version of the network to pre- dict traces from ultrasound only. However hand-tracing the entire dataset of ultrasound images is extremely labor intensive. More- over, the dataset is highly imbalanced since many images are extremely similar. Here we present a bootstrapping method which takes advantage of this imbalance, iteratively se- lecting a small subset of images to be hand- traced, then (re)training the DBN, making use of an entropy-based diversity measure for the initial selection. With this approach we achieve a three-fold reduction in human time required to trace an entire dataset with human-level accuracy. |
Persistent Identifier | http://hdl.handle.net/10722/205582 |
DC Field | Value | Language |
---|---|---|
dc.contributor.author | Berry, J | - |
dc.contributor.author | Fasel, I | - |
dc.contributor.author | Fadiga, L | - |
dc.contributor.author | Archangeli, DB | - |
dc.date.accessioned | 2014-09-20T04:13:58Z | - |
dc.date.available | 2014-09-20T04:13:58Z | - |
dc.date.issued | 2011 | - |
dc.identifier.citation | The 28th International Conference on Machine Learning (ICML 2011), Bellevue, WA., 28 June-2 July 2011. | - |
dc.identifier.uri | http://hdl.handle.net/10722/205582 | - |
dc.description.abstract | Training deep belief networks (DBNs) is nor- mally done with large data sets. In this work, our goal is to predict traces of the surface of the tongue in ultrasound images of the mouth during speech. Performance on this task can be dramatically enhanced by pre- training a DBN jointly on human-supplied traces and ultrasound images, then training a modified version of the network to pre- dict traces from ultrasound only. However hand-tracing the entire dataset of ultrasound images is extremely labor intensive. More- over, the dataset is highly imbalanced since many images are extremely similar. Here we present a bootstrapping method which takes advantage of this imbalance, iteratively se- lecting a small subset of images to be hand- traced, then (re)training the DBN, making use of an entropy-based diversity measure for the initial selection. With this approach we achieve a three-fold reduction in human time required to trace an entire dataset with human-level accuracy. | - |
dc.language | eng | - |
dc.relation.ispartof | International Conference on Machine Learning, ICML 2011 | - |
dc.title | Entropy coding for training deep belief networks with imbalanced and unlabeled data | - |
dc.type | Conference_Paper | - |
dc.identifier.email | Archangeli, DB: darchang@hku.hk | - |
dc.identifier.authority | Archangeli, DB=rp01748 | - |
dc.identifier.hkuros | 232525 | - |